Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 1 | /*===---- avx512vlintrin.h - AVX512VL intrinsics ---------------------------=== |
| 2 | * |
| 3 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 4 | * of this software and associated documentation files (the "Software"), to deal |
| 5 | * in the Software without restriction, including without limitation the rights |
| 6 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 7 | * copies of the Software, and to permit persons to whom the Software is |
| 8 | * furnished to do so, subject to the following conditions: |
| 9 | * |
| 10 | * The above copyright notice and this permission notice shall be included in |
| 11 | * all copies or substantial portions of the Software. |
| 12 | * |
| 13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 15 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 16 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 17 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 18 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 19 | * THE SOFTWARE. |
| 20 | * |
| 21 | *===-----------------------------------------------------------------------=== |
| 22 | */ |
| 23 | |
| 24 | #ifndef __IMMINTRIN_H |
| 25 | #error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead." |
| 26 | #endif |
| 27 | |
| 28 | #ifndef __AVX512VLINTRIN_H |
| 29 | #define __AVX512VLINTRIN_H |
| 30 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 31 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"))) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 32 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 33 | /* Integer compare */ |
| 34 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 35 | #define _mm_cmpeq_epi32_mask(A, B) \ |
| 36 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_EQ) |
| 37 | #define _mm_mask_cmpeq_epi32_mask(k, A, B) \ |
| 38 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 39 | #define _mm_cmpge_epi32_mask(A, B) \ |
| 40 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_GE) |
| 41 | #define _mm_mask_cmpge_epi32_mask(k, A, B) \ |
| 42 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_GE) |
| 43 | #define _mm_cmpgt_epi32_mask(A, B) \ |
| 44 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_GT) |
| 45 | #define _mm_mask_cmpgt_epi32_mask(k, A, B) \ |
| 46 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_GT) |
| 47 | #define _mm_cmple_epi32_mask(A, B) \ |
| 48 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_LE) |
| 49 | #define _mm_mask_cmple_epi32_mask(k, A, B) \ |
| 50 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_LE) |
| 51 | #define _mm_cmplt_epi32_mask(A, B) \ |
| 52 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_LT) |
| 53 | #define _mm_mask_cmplt_epi32_mask(k, A, B) \ |
| 54 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_LT) |
| 55 | #define _mm_cmpneq_epi32_mask(A, B) \ |
| 56 | _mm_cmp_epi32_mask((A), (B), _MM_CMPINT_NE) |
| 57 | #define _mm_mask_cmpneq_epi32_mask(k, A, B) \ |
| 58 | _mm_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_NE) |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 59 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 60 | #define _mm256_cmpeq_epi32_mask(A, B) \ |
| 61 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_EQ) |
| 62 | #define _mm256_mask_cmpeq_epi32_mask(k, A, B) \ |
| 63 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 64 | #define _mm256_cmpge_epi32_mask(A, B) \ |
| 65 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_GE) |
| 66 | #define _mm256_mask_cmpge_epi32_mask(k, A, B) \ |
| 67 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_GE) |
| 68 | #define _mm256_cmpgt_epi32_mask(A, B) \ |
| 69 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_GT) |
| 70 | #define _mm256_mask_cmpgt_epi32_mask(k, A, B) \ |
| 71 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_GT) |
| 72 | #define _mm256_cmple_epi32_mask(A, B) \ |
| 73 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_LE) |
| 74 | #define _mm256_mask_cmple_epi32_mask(k, A, B) \ |
| 75 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_LE) |
| 76 | #define _mm256_cmplt_epi32_mask(A, B) \ |
| 77 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_LT) |
| 78 | #define _mm256_mask_cmplt_epi32_mask(k, A, B) \ |
| 79 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_LT) |
| 80 | #define _mm256_cmpneq_epi32_mask(A, B) \ |
| 81 | _mm256_cmp_epi32_mask((A), (B), _MM_CMPINT_NE) |
| 82 | #define _mm256_mask_cmpneq_epi32_mask(k, A, B) \ |
| 83 | _mm256_mask_cmp_epi32_mask((k), (A), (B), _MM_CMPINT_NE) |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 84 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 85 | #define _mm_cmpeq_epu32_mask(A, B) \ |
| 86 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_EQ) |
| 87 | #define _mm_mask_cmpeq_epu32_mask(k, A, B) \ |
| 88 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 89 | #define _mm_cmpge_epu32_mask(A, B) \ |
| 90 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_GE) |
| 91 | #define _mm_mask_cmpge_epu32_mask(k, A, B) \ |
| 92 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_GE) |
| 93 | #define _mm_cmpgt_epu32_mask(A, B) \ |
| 94 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_GT) |
| 95 | #define _mm_mask_cmpgt_epu32_mask(k, A, B) \ |
| 96 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_GT) |
| 97 | #define _mm_cmple_epu32_mask(A, B) \ |
| 98 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_LE) |
| 99 | #define _mm_mask_cmple_epu32_mask(k, A, B) \ |
| 100 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_LE) |
| 101 | #define _mm_cmplt_epu32_mask(A, B) \ |
| 102 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_LT) |
| 103 | #define _mm_mask_cmplt_epu32_mask(k, A, B) \ |
| 104 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_LT) |
| 105 | #define _mm_cmpneq_epu32_mask(A, B) \ |
| 106 | _mm_cmp_epu32_mask((A), (B), _MM_CMPINT_NE) |
| 107 | #define _mm_mask_cmpneq_epu32_mask(k, A, B) \ |
| 108 | _mm_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_NE) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 109 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 110 | #define _mm256_cmpeq_epu32_mask(A, B) \ |
| 111 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_EQ) |
| 112 | #define _mm256_mask_cmpeq_epu32_mask(k, A, B) \ |
| 113 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 114 | #define _mm256_cmpge_epu32_mask(A, B) \ |
| 115 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_GE) |
| 116 | #define _mm256_mask_cmpge_epu32_mask(k, A, B) \ |
| 117 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_GE) |
| 118 | #define _mm256_cmpgt_epu32_mask(A, B) \ |
| 119 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_GT) |
| 120 | #define _mm256_mask_cmpgt_epu32_mask(k, A, B) \ |
| 121 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_GT) |
| 122 | #define _mm256_cmple_epu32_mask(A, B) \ |
| 123 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_LE) |
| 124 | #define _mm256_mask_cmple_epu32_mask(k, A, B) \ |
| 125 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_LE) |
| 126 | #define _mm256_cmplt_epu32_mask(A, B) \ |
| 127 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_LT) |
| 128 | #define _mm256_mask_cmplt_epu32_mask(k, A, B) \ |
| 129 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_LT) |
| 130 | #define _mm256_cmpneq_epu32_mask(A, B) \ |
| 131 | _mm256_cmp_epu32_mask((A), (B), _MM_CMPINT_NE) |
| 132 | #define _mm256_mask_cmpneq_epu32_mask(k, A, B) \ |
| 133 | _mm256_mask_cmp_epu32_mask((k), (A), (B), _MM_CMPINT_NE) |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 134 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 135 | #define _mm_cmpeq_epi64_mask(A, B) \ |
| 136 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_EQ) |
| 137 | #define _mm_mask_cmpeq_epi64_mask(k, A, B) \ |
| 138 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 139 | #define _mm_cmpge_epi64_mask(A, B) \ |
| 140 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_GE) |
| 141 | #define _mm_mask_cmpge_epi64_mask(k, A, B) \ |
| 142 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_GE) |
| 143 | #define _mm_cmpgt_epi64_mask(A, B) \ |
| 144 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_GT) |
| 145 | #define _mm_mask_cmpgt_epi64_mask(k, A, B) \ |
| 146 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_GT) |
| 147 | #define _mm_cmple_epi64_mask(A, B) \ |
| 148 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_LE) |
| 149 | #define _mm_mask_cmple_epi64_mask(k, A, B) \ |
| 150 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_LE) |
| 151 | #define _mm_cmplt_epi64_mask(A, B) \ |
| 152 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_LT) |
| 153 | #define _mm_mask_cmplt_epi64_mask(k, A, B) \ |
| 154 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_LT) |
| 155 | #define _mm_cmpneq_epi64_mask(A, B) \ |
| 156 | _mm_cmp_epi64_mask((A), (B), _MM_CMPINT_NE) |
| 157 | #define _mm_mask_cmpneq_epi64_mask(k, A, B) \ |
| 158 | _mm_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_NE) |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 159 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 160 | #define _mm256_cmpeq_epi64_mask(A, B) \ |
| 161 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_EQ) |
| 162 | #define _mm256_mask_cmpeq_epi64_mask(k, A, B) \ |
| 163 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 164 | #define _mm256_cmpge_epi64_mask(A, B) \ |
| 165 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_GE) |
| 166 | #define _mm256_mask_cmpge_epi64_mask(k, A, B) \ |
| 167 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_GE) |
| 168 | #define _mm256_cmpgt_epi64_mask(A, B) \ |
| 169 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_GT) |
| 170 | #define _mm256_mask_cmpgt_epi64_mask(k, A, B) \ |
| 171 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_GT) |
| 172 | #define _mm256_cmple_epi64_mask(A, B) \ |
| 173 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_LE) |
| 174 | #define _mm256_mask_cmple_epi64_mask(k, A, B) \ |
| 175 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_LE) |
| 176 | #define _mm256_cmplt_epi64_mask(A, B) \ |
| 177 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_LT) |
| 178 | #define _mm256_mask_cmplt_epi64_mask(k, A, B) \ |
| 179 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_LT) |
| 180 | #define _mm256_cmpneq_epi64_mask(A, B) \ |
| 181 | _mm256_cmp_epi64_mask((A), (B), _MM_CMPINT_NE) |
| 182 | #define _mm256_mask_cmpneq_epi64_mask(k, A, B) \ |
| 183 | _mm256_mask_cmp_epi64_mask((k), (A), (B), _MM_CMPINT_NE) |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 184 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 185 | #define _mm_cmpeq_epu64_mask(A, B) \ |
| 186 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_EQ) |
| 187 | #define _mm_mask_cmpeq_epu64_mask(k, A, B) \ |
| 188 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 189 | #define _mm_cmpge_epu64_mask(A, B) \ |
| 190 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_GE) |
| 191 | #define _mm_mask_cmpge_epu64_mask(k, A, B) \ |
| 192 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_GE) |
| 193 | #define _mm_cmpgt_epu64_mask(A, B) \ |
| 194 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_GT) |
| 195 | #define _mm_mask_cmpgt_epu64_mask(k, A, B) \ |
| 196 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_GT) |
| 197 | #define _mm_cmple_epu64_mask(A, B) \ |
| 198 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_LE) |
| 199 | #define _mm_mask_cmple_epu64_mask(k, A, B) \ |
| 200 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_LE) |
| 201 | #define _mm_cmplt_epu64_mask(A, B) \ |
| 202 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_LT) |
| 203 | #define _mm_mask_cmplt_epu64_mask(k, A, B) \ |
| 204 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_LT) |
| 205 | #define _mm_cmpneq_epu64_mask(A, B) \ |
| 206 | _mm_cmp_epu64_mask((A), (B), _MM_CMPINT_NE) |
| 207 | #define _mm_mask_cmpneq_epu64_mask(k, A, B) \ |
| 208 | _mm_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_NE) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 209 | |
Craig Topper | 57f96ac | 2017-11-06 21:00:49 +0000 | [diff] [blame] | 210 | #define _mm256_cmpeq_epu64_mask(A, B) \ |
| 211 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_EQ) |
| 212 | #define _mm256_mask_cmpeq_epu64_mask(k, A, B) \ |
| 213 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_EQ) |
| 214 | #define _mm256_cmpge_epu64_mask(A, B) \ |
| 215 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_GE) |
| 216 | #define _mm256_mask_cmpge_epu64_mask(k, A, B) \ |
| 217 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_GE) |
| 218 | #define _mm256_cmpgt_epu64_mask(A, B) \ |
| 219 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_GT) |
| 220 | #define _mm256_mask_cmpgt_epu64_mask(k, A, B) \ |
| 221 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_GT) |
| 222 | #define _mm256_cmple_epu64_mask(A, B) \ |
| 223 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_LE) |
| 224 | #define _mm256_mask_cmple_epu64_mask(k, A, B) \ |
| 225 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_LE) |
| 226 | #define _mm256_cmplt_epu64_mask(A, B) \ |
| 227 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_LT) |
| 228 | #define _mm256_mask_cmplt_epu64_mask(k, A, B) \ |
| 229 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_LT) |
| 230 | #define _mm256_cmpneq_epu64_mask(A, B) \ |
| 231 | _mm256_cmp_epu64_mask((A), (B), _MM_CMPINT_NE) |
| 232 | #define _mm256_mask_cmpneq_epu64_mask(k, A, B) \ |
| 233 | _mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_NE) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 234 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 235 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 236 | _mm256_mask_add_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 237 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 238 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 239 | (__v8si)_mm256_add_epi32(__A, __B), |
| 240 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 241 | } |
| 242 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 243 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 244 | _mm256_maskz_add_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 245 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 246 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 247 | (__v8si)_mm256_add_epi32(__A, __B), |
| 248 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 249 | } |
| 250 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 251 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 252 | _mm256_mask_add_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 253 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 254 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 255 | (__v4di)_mm256_add_epi64(__A, __B), |
| 256 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 257 | } |
| 258 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 259 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 260 | _mm256_maskz_add_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 261 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 262 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 263 | (__v4di)_mm256_add_epi64(__A, __B), |
| 264 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 265 | } |
| 266 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 267 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 268 | _mm256_mask_sub_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 269 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 270 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 271 | (__v8si)_mm256_sub_epi32(__A, __B), |
| 272 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 273 | } |
| 274 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 275 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 276 | _mm256_maskz_sub_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 277 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 278 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 279 | (__v8si)_mm256_sub_epi32(__A, __B), |
| 280 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 281 | } |
| 282 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 283 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 284 | _mm256_mask_sub_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 285 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 286 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 287 | (__v4di)_mm256_sub_epi64(__A, __B), |
| 288 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 289 | } |
| 290 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 291 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 292 | _mm256_maskz_sub_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 293 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 294 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 295 | (__v4di)_mm256_sub_epi64(__A, __B), |
| 296 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 297 | } |
| 298 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 299 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 300 | _mm_mask_add_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 301 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 302 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 303 | (__v4si)_mm_add_epi32(__A, __B), |
| 304 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 305 | } |
| 306 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 307 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 308 | _mm_maskz_add_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 309 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 310 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 311 | (__v4si)_mm_add_epi32(__A, __B), |
| 312 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 313 | } |
| 314 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 315 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 316 | _mm_mask_add_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 317 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 318 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 319 | (__v2di)_mm_add_epi64(__A, __B), |
| 320 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 321 | } |
| 322 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 323 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 324 | _mm_maskz_add_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 325 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 326 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 327 | (__v2di)_mm_add_epi64(__A, __B), |
| 328 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 329 | } |
| 330 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 331 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 332 | _mm_mask_sub_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 333 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 334 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 335 | (__v4si)_mm_sub_epi32(__A, __B), |
| 336 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 337 | } |
| 338 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 339 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 340 | _mm_maskz_sub_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 341 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 342 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 343 | (__v4si)_mm_sub_epi32(__A, __B), |
| 344 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 345 | } |
| 346 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 347 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 348 | _mm_mask_sub_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 349 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 350 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 351 | (__v2di)_mm_sub_epi64(__A, __B), |
| 352 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 353 | } |
| 354 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 355 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 356 | _mm_maskz_sub_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 357 | { |
Craig Topper | 0e18976 | 2016-09-03 18:29:35 +0000 | [diff] [blame] | 358 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 359 | (__v2di)_mm_sub_epi64(__A, __B), |
| 360 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 361 | } |
| 362 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 363 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 364 | _mm256_mask_mul_epi32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 365 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 366 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 367 | (__v4di)_mm256_mul_epi32(__X, __Y), |
| 368 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 369 | } |
| 370 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 371 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 372 | _mm256_maskz_mul_epi32(__mmask8 __M, __m256i __X, __m256i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 373 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 374 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 375 | (__v4di)_mm256_mul_epi32(__X, __Y), |
| 376 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 377 | } |
| 378 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 379 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 380 | _mm_mask_mul_epi32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 381 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 382 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 383 | (__v2di)_mm_mul_epi32(__X, __Y), |
| 384 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 385 | } |
| 386 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 387 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 388 | _mm_maskz_mul_epi32(__mmask8 __M, __m128i __X, __m128i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 389 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 390 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 391 | (__v2di)_mm_mul_epi32(__X, __Y), |
| 392 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 393 | } |
| 394 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 395 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 396 | _mm256_mask_mul_epu32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 397 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 398 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 399 | (__v4di)_mm256_mul_epu32(__X, __Y), |
| 400 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 401 | } |
| 402 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 403 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 404 | _mm256_maskz_mul_epu32(__mmask8 __M, __m256i __X, __m256i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 405 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 406 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 407 | (__v4di)_mm256_mul_epu32(__X, __Y), |
| 408 | (__v4di)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 409 | } |
| 410 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 411 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 412 | _mm_mask_mul_epu32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 413 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 414 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 415 | (__v2di)_mm_mul_epu32(__X, __Y), |
| 416 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 417 | } |
| 418 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 419 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 420 | _mm_maskz_mul_epu32(__mmask8 __M, __m128i __X, __m128i __Y) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 421 | { |
Craig Topper | 09e9400 | 2016-10-29 19:02:07 +0000 | [diff] [blame] | 422 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 423 | (__v2di)_mm_mul_epu32(__X, __Y), |
| 424 | (__v2di)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 425 | } |
| 426 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 427 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 428 | _mm256_maskz_mullo_epi32(__mmask8 __M, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 429 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 430 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 431 | (__v8si)_mm256_mullo_epi32(__A, __B), |
| 432 | (__v8si)_mm256_setzero_si256()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 433 | } |
| 434 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 435 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 436 | _mm256_mask_mullo_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 437 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 438 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 439 | (__v8si)_mm256_mullo_epi32(__A, __B), |
| 440 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 441 | } |
| 442 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 443 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 444 | _mm_maskz_mullo_epi32(__mmask8 __M, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 445 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 446 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 447 | (__v4si)_mm_mullo_epi32(__A, __B), |
| 448 | (__v4si)_mm_setzero_si128()); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 449 | } |
| 450 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 451 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 452 | _mm_mask_mullo_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 453 | { |
Craig Topper | f43e4a1 | 2016-09-03 19:19:49 +0000 | [diff] [blame] | 454 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 455 | (__v4si)_mm_mullo_epi32(__A, __B), |
| 456 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 457 | } |
| 458 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 459 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 460 | _mm256_mask_and_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 461 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 462 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 463 | (__v8si)_mm256_and_si256(__A, __B), |
| 464 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 465 | } |
| 466 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 467 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 468 | _mm256_maskz_and_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 469 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 470 | return (__m256i)_mm256_mask_and_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 471 | } |
| 472 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 473 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 474 | _mm_mask_and_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 475 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 476 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 477 | (__v4si)_mm_and_si128(__A, __B), |
| 478 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 479 | } |
| 480 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 481 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 482 | _mm_maskz_and_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 483 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 484 | return (__m128i)_mm_mask_and_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 485 | } |
| 486 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 487 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 488 | _mm256_mask_andnot_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 489 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 490 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 491 | (__v8si)_mm256_andnot_si256(__A, __B), |
| 492 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 493 | } |
| 494 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 495 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 496 | _mm256_maskz_andnot_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 497 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 498 | return (__m256i)_mm256_mask_andnot_epi32(_mm256_setzero_si256(), |
| 499 | __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 500 | } |
| 501 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 502 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 503 | _mm_mask_andnot_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 504 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 505 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 506 | (__v4si)_mm_andnot_si128(__A, __B), |
| 507 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 508 | } |
| 509 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 510 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 511 | _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 512 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 513 | return (__m128i)_mm_mask_andnot_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 514 | } |
| 515 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 516 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 517 | _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 518 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 519 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 520 | (__v8si)_mm256_or_si256(__A, __B), |
| 521 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 522 | } |
| 523 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 524 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 525 | _mm256_maskz_or_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 526 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 527 | return (__m256i)_mm256_mask_or_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 528 | } |
| 529 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 530 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 531 | _mm_mask_or_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 532 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 533 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 534 | (__v4si)_mm_or_si128(__A, __B), |
| 535 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 536 | } |
| 537 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 538 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 539 | _mm_maskz_or_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 540 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 541 | return (__m128i)_mm_mask_or_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 542 | } |
| 543 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 544 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 545 | _mm256_mask_xor_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 546 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 547 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 548 | (__v8si)_mm256_xor_si256(__A, __B), |
| 549 | (__v8si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 550 | } |
| 551 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 552 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 553 | _mm256_maskz_xor_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 554 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 555 | return (__m256i)_mm256_mask_xor_epi32(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 556 | } |
| 557 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 558 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 559 | _mm_mask_xor_epi32(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 560 | __m128i __B) |
| 561 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 562 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 563 | (__v4si)_mm_xor_si128(__A, __B), |
| 564 | (__v4si)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 565 | } |
| 566 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 567 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 568 | _mm_maskz_xor_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 569 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 570 | return (__m128i)_mm_mask_xor_epi32(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 571 | } |
| 572 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 573 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 574 | _mm256_mask_and_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 575 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 576 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 577 | (__v4di)_mm256_and_si256(__A, __B), |
| 578 | (__v4di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 579 | } |
| 580 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 581 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 582 | _mm256_maskz_and_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 583 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 584 | return (__m256i)_mm256_mask_and_epi64(_mm256_setzero_si256(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 585 | } |
| 586 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 587 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 588 | _mm_mask_and_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 589 | { |
| 590 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 591 | (__v2di)_mm_and_si128(__A, __B), |
| 592 | (__v2di)__W); |
| 593 | } |
| 594 | |
| 595 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 596 | _mm_maskz_and_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 597 | { |
| 598 | return (__m128i)_mm_mask_and_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 599 | } |
| 600 | |
| 601 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 602 | _mm256_mask_andnot_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 603 | { |
| 604 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 605 | (__v4di)_mm256_andnot_si256(__A, __B), |
| 606 | (__v4di)__W); |
| 607 | } |
| 608 | |
| 609 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 610 | _mm256_maskz_andnot_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 611 | { |
| 612 | return (__m256i)_mm256_mask_andnot_epi64(_mm256_setzero_si256(), |
| 613 | __U, __A, __B); |
| 614 | } |
| 615 | |
| 616 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 617 | _mm_mask_andnot_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 618 | { |
| 619 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 620 | (__v2di)_mm_andnot_si128(__A, __B), |
| 621 | (__v2di)__W); |
| 622 | } |
| 623 | |
| 624 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 625 | _mm_maskz_andnot_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 626 | { |
| 627 | return (__m128i)_mm_mask_andnot_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 628 | } |
| 629 | |
| 630 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 631 | _mm256_mask_or_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 632 | { |
| 633 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 634 | (__v4di)_mm256_or_si256(__A, __B), |
| 635 | (__v4di)__W); |
| 636 | } |
| 637 | |
| 638 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 639 | _mm256_maskz_or_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 640 | { |
| 641 | return (__m256i)_mm256_mask_or_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 642 | } |
| 643 | |
| 644 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 645 | _mm_mask_or_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 646 | { |
| 647 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 648 | (__v2di)_mm_or_si128(__A, __B), |
| 649 | (__v2di)__W); |
| 650 | } |
| 651 | |
| 652 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 653 | _mm_maskz_or_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 654 | { |
| 655 | return (__m128i)_mm_mask_or_epi64(_mm_setzero_si128(), __U, __A, __B); |
| 656 | } |
| 657 | |
| 658 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 659 | _mm256_mask_xor_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
| 660 | { |
| 661 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 662 | (__v4di)_mm256_xor_si256(__A, __B), |
| 663 | (__v4di)__W); |
| 664 | } |
| 665 | |
| 666 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 667 | _mm256_maskz_xor_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
| 668 | { |
| 669 | return (__m256i)_mm256_mask_xor_epi64(_mm256_setzero_si256(), __U, __A, __B); |
| 670 | } |
| 671 | |
| 672 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 673 | _mm_mask_xor_epi64(__m128i __W, __mmask8 __U, __m128i __A, |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 674 | __m128i __B) |
| 675 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 676 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 677 | (__v2di)_mm_xor_si128(__A, __B), |
| 678 | (__v2di)__W); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 679 | } |
| 680 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 681 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 682 | _mm_maskz_xor_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 683 | { |
Craig Topper | 4d61a3c | 2016-07-11 06:14:18 +0000 | [diff] [blame] | 684 | return (__m128i)_mm_mask_xor_epi64(_mm_setzero_si128(), __U, __A, __B); |
Elena Demikhovsky | 35dc8c0 | 2015-04-28 13:28:01 +0000 | [diff] [blame] | 685 | } |
| 686 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 687 | #define _mm_cmp_epi32_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 688 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 689 | (__v4si)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 690 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 691 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 692 | #define _mm_mask_cmp_epi32_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 693 | (__mmask8)__builtin_ia32_cmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 694 | (__v4si)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 695 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 696 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 697 | #define _mm_cmp_epu32_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 698 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 699 | (__v4si)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 700 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 701 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 702 | #define _mm_mask_cmp_epu32_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 703 | (__mmask8)__builtin_ia32_ucmpd128_mask((__v4si)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 704 | (__v4si)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 705 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 706 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 707 | #define _mm256_cmp_epi32_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 708 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 709 | (__v8si)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 710 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 711 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 712 | #define _mm256_mask_cmp_epi32_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 713 | (__mmask8)__builtin_ia32_cmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 714 | (__v8si)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 715 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 716 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 717 | #define _mm256_cmp_epu32_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 718 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 719 | (__v8si)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 720 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 721 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 722 | #define _mm256_mask_cmp_epu32_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 723 | (__mmask8)__builtin_ia32_ucmpd256_mask((__v8si)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 724 | (__v8si)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 725 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 726 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 727 | #define _mm_cmp_epi64_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 728 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 729 | (__v2di)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 730 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 731 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 732 | #define _mm_mask_cmp_epi64_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 733 | (__mmask8)__builtin_ia32_cmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 734 | (__v2di)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 735 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 736 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 737 | #define _mm_cmp_epu64_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 738 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 739 | (__v2di)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 740 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 741 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 742 | #define _mm_mask_cmp_epu64_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 743 | (__mmask8)__builtin_ia32_ucmpq128_mask((__v2di)(__m128i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 744 | (__v2di)(__m128i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 745 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 746 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 747 | #define _mm256_cmp_epi64_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 748 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 749 | (__v4di)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 750 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 751 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 752 | #define _mm256_mask_cmp_epi64_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 753 | (__mmask8)__builtin_ia32_cmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 754 | (__v4di)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 755 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 756 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 757 | #define _mm256_cmp_epu64_mask(a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 758 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 759 | (__v4di)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 760 | (__mmask8)-1) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 761 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 762 | #define _mm256_mask_cmp_epu64_mask(m, a, b, p) \ |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 763 | (__mmask8)__builtin_ia32_ucmpq256_mask((__v4di)(__m256i)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 764 | (__v4di)(__m256i)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 765 | (__mmask8)(m)) |
Craig Topper | 2f25a5a | 2015-01-26 08:11:49 +0000 | [diff] [blame] | 766 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 767 | #define _mm256_cmp_ps_mask(a, b, p) \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 768 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 769 | (__v8sf)(__m256)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 770 | (__mmask8)-1) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 771 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 772 | #define _mm256_mask_cmp_ps_mask(m, a, b, p) \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 773 | (__mmask8)__builtin_ia32_cmpps256_mask((__v8sf)(__m256)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 774 | (__v8sf)(__m256)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 775 | (__mmask8)(m)) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 776 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 777 | #define _mm256_cmp_pd_mask(a, b, p) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 778 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 779 | (__v4df)(__m256d)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 780 | (__mmask8)-1) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 781 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 782 | #define _mm256_mask_cmp_pd_mask(m, a, b, p) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 783 | (__mmask8)__builtin_ia32_cmppd256_mask((__v4df)(__m256d)(a), \ |
| 784 | (__v4df)(__m256d)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 785 | (__mmask8)(m)) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 786 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 787 | #define _mm_cmp_ps_mask(a, b, p) \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 788 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 789 | (__v4sf)(__m128)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 790 | (__mmask8)-1) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 791 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 792 | #define _mm_mask_cmp_ps_mask(m, a, b, p) \ |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 793 | (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 794 | (__v4sf)(__m128)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 795 | (__mmask8)(m)) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 796 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 797 | #define _mm_cmp_pd_mask(a, b, p) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 798 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 799 | (__v2df)(__m128d)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 800 | (__mmask8)-1) |
Elena Demikhovsky | bd5c8b9 | 2015-05-07 11:26:36 +0000 | [diff] [blame] | 801 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 802 | #define _mm_mask_cmp_pd_mask(m, a, b, p) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 803 | (__mmask8)__builtin_ia32_cmppd128_mask((__v2df)(__m128d)(a), \ |
| 804 | (__v2df)(__m128d)(b), (int)(p), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 805 | (__mmask8)(m)) |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 806 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 807 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 808 | _mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 809 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 810 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 811 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 812 | (__v2df) __B, |
| 813 | (__v2df) __C), |
| 814 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 815 | } |
| 816 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 817 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 818 | _mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 819 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 820 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 821 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 822 | (__v2df) __B, |
| 823 | (__v2df) __C), |
| 824 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 825 | } |
| 826 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 827 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 828 | _mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 829 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 830 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 831 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 832 | (__v2df) __B, |
| 833 | (__v2df) __C), |
| 834 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 835 | } |
| 836 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 837 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 838 | _mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 839 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 840 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 841 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 842 | (__v2df) __B, |
| 843 | -(__v2df) __C), |
| 844 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 845 | } |
| 846 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 847 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 848 | _mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 849 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 850 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 851 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 852 | (__v2df) __B, |
| 853 | -(__v2df) __C), |
| 854 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 855 | } |
| 856 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 857 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 858 | _mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 859 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 860 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 861 | __builtin_ia32_vfmaddpd (-(__v2df) __A, |
| 862 | (__v2df) __B, |
| 863 | (__v2df) __C), |
| 864 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 865 | } |
| 866 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 867 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 868 | _mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 869 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 870 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 871 | __builtin_ia32_vfmaddpd (-(__v2df) __A, |
| 872 | (__v2df) __B, |
| 873 | (__v2df) __C), |
| 874 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 875 | } |
| 876 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 877 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 878 | _mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 879 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 880 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 881 | __builtin_ia32_vfmaddpd (-(__v2df) __A, |
| 882 | (__v2df) __B, |
| 883 | -(__v2df) __C), |
| 884 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 885 | } |
| 886 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 887 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 888 | _mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 889 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 890 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 891 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 892 | (__v4df) __B, |
| 893 | (__v4df) __C), |
| 894 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 895 | } |
| 896 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 897 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 898 | _mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 899 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 900 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 901 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 902 | (__v4df) __B, |
| 903 | (__v4df) __C), |
| 904 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 905 | } |
| 906 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 907 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 908 | _mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 909 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 910 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 911 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 912 | (__v4df) __B, |
| 913 | (__v4df) __C), |
| 914 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 915 | } |
| 916 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 917 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 918 | _mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 919 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 920 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 921 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 922 | (__v4df) __B, |
| 923 | -(__v4df) __C), |
| 924 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 925 | } |
| 926 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 927 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 928 | _mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 929 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 930 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 931 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 932 | (__v4df) __B, |
| 933 | -(__v4df) __C), |
| 934 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 935 | } |
| 936 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 937 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 938 | _mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 939 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 940 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 941 | __builtin_ia32_vfmaddpd256 (-(__v4df) __A, |
| 942 | (__v4df) __B, |
| 943 | (__v4df) __C), |
| 944 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 945 | } |
| 946 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 947 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 948 | _mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 949 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 950 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 951 | __builtin_ia32_vfmaddpd256 (-(__v4df) __A, |
| 952 | (__v4df) __B, |
| 953 | (__v4df) __C), |
| 954 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 955 | } |
| 956 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 957 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 958 | _mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 959 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 960 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 961 | __builtin_ia32_vfmaddpd256 (-(__v4df) __A, |
| 962 | (__v4df) __B, |
| 963 | -(__v4df) __C), |
| 964 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 965 | } |
| 966 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 967 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 968 | _mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 969 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 970 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 971 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 972 | (__v4sf) __B, |
| 973 | (__v4sf) __C), |
| 974 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 975 | } |
| 976 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 977 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 978 | _mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 979 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 980 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 981 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 982 | (__v4sf) __B, |
| 983 | (__v4sf) __C), |
| 984 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 985 | } |
| 986 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 987 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 988 | _mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 989 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 990 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 991 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 992 | (__v4sf) __B, |
| 993 | (__v4sf) __C), |
| 994 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 995 | } |
| 996 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 997 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 998 | _mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 999 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1000 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1001 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1002 | (__v4sf) __B, |
| 1003 | -(__v4sf) __C), |
| 1004 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1005 | } |
| 1006 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1007 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1008 | _mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1009 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1010 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1011 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1012 | (__v4sf) __B, |
| 1013 | -(__v4sf) __C), |
| 1014 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1015 | } |
| 1016 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1017 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1018 | _mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1019 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1020 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1021 | __builtin_ia32_vfmaddps (-(__v4sf) __A, |
| 1022 | (__v4sf) __B, |
| 1023 | (__v4sf) __C), |
| 1024 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1025 | } |
| 1026 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1027 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1028 | _mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1029 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1030 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1031 | __builtin_ia32_vfmaddps (-(__v4sf) __A, |
| 1032 | (__v4sf) __B, |
| 1033 | (__v4sf) __C), |
| 1034 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1035 | } |
| 1036 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1037 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1038 | _mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1039 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1040 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1041 | __builtin_ia32_vfmaddps (-(__v4sf) __A, |
| 1042 | (__v4sf) __B, |
| 1043 | -(__v4sf) __C), |
| 1044 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1045 | } |
| 1046 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1047 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1048 | _mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1049 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1050 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1051 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1052 | (__v8sf) __B, |
| 1053 | (__v8sf) __C), |
| 1054 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1055 | } |
| 1056 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1057 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1058 | _mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1059 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1060 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1061 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1062 | (__v8sf) __B, |
| 1063 | (__v8sf) __C), |
| 1064 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1065 | } |
| 1066 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1067 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1068 | _mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1069 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1070 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1071 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1072 | (__v8sf) __B, |
| 1073 | (__v8sf) __C), |
| 1074 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1075 | } |
| 1076 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1077 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1078 | _mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1079 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1080 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1081 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1082 | (__v8sf) __B, |
| 1083 | -(__v8sf) __C), |
| 1084 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1085 | } |
| 1086 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1087 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1088 | _mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1089 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1090 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1091 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1092 | (__v8sf) __B, |
| 1093 | -(__v8sf) __C), |
| 1094 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1095 | } |
| 1096 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1097 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1098 | _mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1099 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1100 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1101 | __builtin_ia32_vfmaddps256 (-(__v8sf) __A, |
| 1102 | (__v8sf) __B, |
| 1103 | (__v8sf) __C), |
| 1104 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1105 | } |
| 1106 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1107 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1108 | _mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1109 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1110 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1111 | __builtin_ia32_vfmaddps256 (-(__v8sf) __A, |
| 1112 | (__v8sf) __B, |
| 1113 | (__v8sf) __C), |
| 1114 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1115 | } |
| 1116 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1117 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1118 | _mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1119 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1120 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1121 | __builtin_ia32_vfmaddps256 (-(__v8sf) __A, |
| 1122 | (__v8sf) __B, |
| 1123 | -(__v8sf) __C), |
| 1124 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1125 | } |
| 1126 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1127 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1128 | _mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1129 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1130 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1131 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1132 | (__v2df) __B, |
| 1133 | (__v2df) __C), |
| 1134 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1135 | } |
| 1136 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1137 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1138 | _mm_mask3_fmaddsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1139 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1140 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1141 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1142 | (__v2df) __B, |
| 1143 | (__v2df) __C), |
| 1144 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1145 | } |
| 1146 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1147 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1148 | _mm_maskz_fmaddsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1149 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1150 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1151 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1152 | (__v2df) __B, |
| 1153 | (__v2df) __C), |
| 1154 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1155 | } |
| 1156 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1157 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1158 | _mm_mask_fmsubadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1159 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1160 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1161 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1162 | (__v2df) __B, |
| 1163 | -(__v2df) __C), |
| 1164 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1165 | } |
| 1166 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1167 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1168 | _mm_maskz_fmsubadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) |
| 1169 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1170 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1171 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1172 | (__v2df) __B, |
| 1173 | -(__v2df) __C), |
| 1174 | (__v2df)_mm_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1175 | } |
| 1176 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1177 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1178 | _mm256_mask_fmaddsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1179 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1180 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1181 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1182 | (__v4df) __B, |
| 1183 | (__v4df) __C), |
| 1184 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1185 | } |
| 1186 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1187 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1188 | _mm256_mask3_fmaddsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1189 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1190 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1191 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1192 | (__v4df) __B, |
| 1193 | (__v4df) __C), |
| 1194 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1195 | } |
| 1196 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1197 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1198 | _mm256_maskz_fmaddsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1199 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1200 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1201 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1202 | (__v4df) __B, |
| 1203 | (__v4df) __C), |
| 1204 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1205 | } |
| 1206 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1207 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1208 | _mm256_mask_fmsubadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1209 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1210 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1211 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1212 | (__v4df) __B, |
| 1213 | -(__v4df) __C), |
| 1214 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1215 | } |
| 1216 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1217 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1218 | _mm256_maskz_fmsubadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) |
| 1219 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1220 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1221 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1222 | (__v4df) __B, |
| 1223 | -(__v4df) __C), |
| 1224 | (__v4df)_mm256_setzero_pd()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1225 | } |
| 1226 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1227 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1228 | _mm_mask_fmaddsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1229 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1230 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1231 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1232 | (__v4sf) __B, |
| 1233 | (__v4sf) __C), |
| 1234 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1235 | } |
| 1236 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1237 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1238 | _mm_mask3_fmaddsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1239 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1240 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1241 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1242 | (__v4sf) __B, |
| 1243 | (__v4sf) __C), |
| 1244 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1245 | } |
| 1246 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1247 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1248 | _mm_maskz_fmaddsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1249 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1250 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1251 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1252 | (__v4sf) __B, |
| 1253 | (__v4sf) __C), |
| 1254 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1255 | } |
| 1256 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1257 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1258 | _mm_mask_fmsubadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1259 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1260 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1261 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1262 | (__v4sf) __B, |
| 1263 | -(__v4sf) __C), |
| 1264 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1265 | } |
| 1266 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1267 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1268 | _mm_maskz_fmsubadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) |
| 1269 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1270 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1271 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1272 | (__v4sf) __B, |
| 1273 | -(__v4sf) __C), |
| 1274 | (__v4sf)_mm_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1275 | } |
| 1276 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1277 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1278 | _mm256_mask_fmaddsub_ps(__m256 __A, __mmask8 __U, __m256 __B, |
| 1279 | __m256 __C) |
| 1280 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1281 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1282 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1283 | (__v8sf) __B, |
| 1284 | (__v8sf) __C), |
| 1285 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1286 | } |
| 1287 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1288 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1289 | _mm256_mask3_fmaddsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1290 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1291 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1292 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1293 | (__v8sf) __B, |
| 1294 | (__v8sf) __C), |
| 1295 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1296 | } |
| 1297 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1298 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1299 | _mm256_maskz_fmaddsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1300 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1301 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1302 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1303 | (__v8sf) __B, |
| 1304 | (__v8sf) __C), |
| 1305 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1306 | } |
| 1307 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1308 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1309 | _mm256_mask_fmsubadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1310 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1311 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1312 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1313 | (__v8sf) __B, |
| 1314 | -(__v8sf) __C), |
| 1315 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1316 | } |
| 1317 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1318 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1319 | _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) |
| 1320 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1321 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1322 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1323 | (__v8sf) __B, |
| 1324 | -(__v8sf) __C), |
| 1325 | (__v8sf)_mm256_setzero_ps()); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1326 | } |
| 1327 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1328 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1329 | _mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1330 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1331 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1332 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 1333 | (__v2df) __B, |
| 1334 | -(__v2df) __C), |
| 1335 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1336 | } |
| 1337 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1338 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1339 | _mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1340 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1341 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1342 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 1343 | (__v4df) __B, |
| 1344 | -(__v4df) __C), |
| 1345 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1346 | } |
| 1347 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1348 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1349 | _mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1350 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1351 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1352 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1353 | (__v4sf) __B, |
| 1354 | -(__v4sf) __C), |
| 1355 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1356 | } |
| 1357 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1358 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1359 | _mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1360 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1361 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1362 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1363 | (__v8sf) __B, |
| 1364 | -(__v8sf) __C), |
| 1365 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1366 | } |
| 1367 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1368 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1369 | _mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1370 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1371 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1372 | __builtin_ia32_vfmaddsubpd ((__v2df) __A, |
| 1373 | (__v2df) __B, |
| 1374 | -(__v2df) __C), |
| 1375 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1376 | } |
| 1377 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1378 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1379 | _mm256_mask3_fmsubadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1380 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1381 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1382 | __builtin_ia32_vfmaddsubpd256 ((__v4df) __A, |
| 1383 | (__v4df) __B, |
| 1384 | -(__v4df) __C), |
| 1385 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1386 | } |
| 1387 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1388 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1389 | _mm_mask3_fmsubadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1390 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1391 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1392 | __builtin_ia32_vfmaddsubps ((__v4sf) __A, |
| 1393 | (__v4sf) __B, |
| 1394 | -(__v4sf) __C), |
| 1395 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1396 | } |
| 1397 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1398 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1399 | _mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1400 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1401 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1402 | __builtin_ia32_vfmaddsubps256 ((__v8sf) __A, |
| 1403 | (__v8sf) __B, |
| 1404 | -(__v8sf) __C), |
| 1405 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1406 | } |
| 1407 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1408 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1409 | _mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1410 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1411 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1412 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 1413 | -(__v2df) __B, |
| 1414 | (__v2df) __C), |
| 1415 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1416 | } |
| 1417 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1418 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1419 | _mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1420 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1421 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1422 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 1423 | -(__v4df) __B, |
| 1424 | (__v4df) __C), |
| 1425 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1426 | } |
| 1427 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1428 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1429 | _mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1430 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1431 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1432 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1433 | -(__v4sf) __B, |
| 1434 | (__v4sf) __C), |
| 1435 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1436 | } |
| 1437 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1438 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1439 | _mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1440 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1441 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1442 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1443 | -(__v8sf) __B, |
| 1444 | (__v8sf) __C), |
| 1445 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1446 | } |
| 1447 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1448 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1449 | _mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) |
| 1450 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1451 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1452 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 1453 | -(__v2df) __B, |
| 1454 | -(__v2df) __C), |
| 1455 | (__v2df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1456 | } |
| 1457 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1458 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1459 | _mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) |
| 1460 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1461 | return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U, |
| 1462 | __builtin_ia32_vfmaddpd ((__v2df) __A, |
| 1463 | -(__v2df) __B, |
| 1464 | -(__v2df) __C), |
| 1465 | (__v2df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1466 | } |
| 1467 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1468 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1469 | _mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) |
| 1470 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1471 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1472 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 1473 | -(__v4df) __B, |
| 1474 | -(__v4df) __C), |
| 1475 | (__v4df) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1476 | } |
| 1477 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1478 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1479 | _mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) |
| 1480 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1481 | return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U, |
| 1482 | __builtin_ia32_vfmaddpd256 ((__v4df) __A, |
| 1483 | -(__v4df) __B, |
| 1484 | -(__v4df) __C), |
| 1485 | (__v4df) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1486 | } |
| 1487 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1488 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1489 | _mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) |
| 1490 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1491 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1492 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1493 | -(__v4sf) __B, |
| 1494 | -(__v4sf) __C), |
| 1495 | (__v4sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1496 | } |
| 1497 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1498 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1499 | _mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) |
| 1500 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1501 | return (__m128) __builtin_ia32_selectps_128((__mmask8) __U, |
| 1502 | __builtin_ia32_vfmaddps ((__v4sf) __A, |
| 1503 | -(__v4sf) __B, |
| 1504 | -(__v4sf) __C), |
| 1505 | (__v4sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1506 | } |
| 1507 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1508 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1509 | _mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) |
| 1510 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1511 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1512 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1513 | -(__v8sf) __B, |
| 1514 | -(__v8sf) __C), |
| 1515 | (__v8sf) __A); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1516 | } |
| 1517 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 1518 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1519 | _mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) |
| 1520 | { |
Gabor Buella | 70d8d51 | 2018-05-30 15:27:49 +0000 | [diff] [blame] | 1521 | return (__m256) __builtin_ia32_selectps_256((__mmask8) __U, |
| 1522 | __builtin_ia32_vfmaddps256 ((__v8sf) __A, |
| 1523 | -(__v8sf) __B, |
| 1524 | -(__v8sf) __C), |
| 1525 | (__v8sf) __C); |
Elena Demikhovsky | c563c2c | 2015-06-29 09:20:57 +0000 | [diff] [blame] | 1526 | } |
| 1527 | |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1528 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1529 | _mm_mask_add_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 1530 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 1531 | (__v2df)_mm_add_pd(__A, __B), |
| 1532 | (__v2df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1533 | } |
| 1534 | |
| 1535 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1536 | _mm_maskz_add_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 1537 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 1538 | (__v2df)_mm_add_pd(__A, __B), |
| 1539 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1540 | } |
| 1541 | |
| 1542 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1543 | _mm256_mask_add_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 1544 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 1545 | (__v4df)_mm256_add_pd(__A, __B), |
| 1546 | (__v4df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1547 | } |
| 1548 | |
| 1549 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1550 | _mm256_maskz_add_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 1551 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 1552 | (__v4df)_mm256_add_pd(__A, __B), |
| 1553 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1554 | } |
| 1555 | |
| 1556 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1557 | _mm_mask_add_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 1558 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1559 | (__v4sf)_mm_add_ps(__A, __B), |
| 1560 | (__v4sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1561 | } |
| 1562 | |
| 1563 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1564 | _mm_maskz_add_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 1565 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1566 | (__v4sf)_mm_add_ps(__A, __B), |
| 1567 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1568 | } |
| 1569 | |
| 1570 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1571 | _mm256_mask_add_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 1572 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1573 | (__v8sf)_mm256_add_ps(__A, __B), |
| 1574 | (__v8sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1575 | } |
| 1576 | |
| 1577 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 1578 | _mm256_maskz_add_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 1579 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1580 | (__v8sf)_mm256_add_ps(__A, __B), |
| 1581 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1582 | } |
| 1583 | |
| 1584 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1585 | _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1586 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1587 | (__v4si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1588 | (__v4si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1589 | } |
| 1590 | |
| 1591 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1592 | _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1593 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1594 | (__v8si) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1595 | (__v8si) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1596 | } |
| 1597 | |
| 1598 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1599 | _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1600 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1601 | (__v2df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1602 | (__v2df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1603 | } |
| 1604 | |
| 1605 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1606 | _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1607 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1608 | (__v4df) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1609 | (__v4df) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1610 | } |
| 1611 | |
| 1612 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1613 | _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1614 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1615 | (__v4sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1616 | (__v4sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1617 | } |
| 1618 | |
| 1619 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1620 | _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1621 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1622 | (__v8sf) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1623 | (__v8sf) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1624 | } |
| 1625 | |
| 1626 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1627 | _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1628 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1629 | (__v2di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1630 | (__v2di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1631 | } |
| 1632 | |
| 1633 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1634 | _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1635 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1636 | (__v4di) __W, |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 1637 | (__v4di) __A); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1638 | } |
| 1639 | |
| 1640 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1641 | _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 1642 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 1643 | (__v2df) __W, |
| 1644 | (__mmask8) __U); |
| 1645 | } |
| 1646 | |
| 1647 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1648 | _mm_maskz_compress_pd (__mmask8 __U, __m128d __A) { |
| 1649 | return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, |
| 1650 | (__v2df) |
| 1651 | _mm_setzero_pd (), |
| 1652 | (__mmask8) __U); |
| 1653 | } |
| 1654 | |
| 1655 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1656 | _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 1657 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 1658 | (__v4df) __W, |
| 1659 | (__mmask8) __U); |
| 1660 | } |
| 1661 | |
| 1662 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1663 | _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) { |
| 1664 | return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, |
| 1665 | (__v4df) |
| 1666 | _mm256_setzero_pd (), |
| 1667 | (__mmask8) __U); |
| 1668 | } |
| 1669 | |
| 1670 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1671 | _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 1672 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 1673 | (__v2di) __W, |
| 1674 | (__mmask8) __U); |
| 1675 | } |
| 1676 | |
| 1677 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1678 | _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) { |
| 1679 | return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, |
| 1680 | (__v2di) |
| 1681 | _mm_setzero_si128 (), |
| 1682 | (__mmask8) __U); |
| 1683 | } |
| 1684 | |
| 1685 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1686 | _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 1687 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 1688 | (__v4di) __W, |
| 1689 | (__mmask8) __U); |
| 1690 | } |
| 1691 | |
| 1692 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1693 | _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) { |
| 1694 | return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, |
| 1695 | (__v4di) |
| 1696 | _mm256_setzero_si256 (), |
| 1697 | (__mmask8) __U); |
| 1698 | } |
| 1699 | |
| 1700 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1701 | _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 1702 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 1703 | (__v4sf) __W, |
| 1704 | (__mmask8) __U); |
| 1705 | } |
| 1706 | |
| 1707 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1708 | _mm_maskz_compress_ps (__mmask8 __U, __m128 __A) { |
| 1709 | return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, |
| 1710 | (__v4sf) |
| 1711 | _mm_setzero_ps (), |
| 1712 | (__mmask8) __U); |
| 1713 | } |
| 1714 | |
| 1715 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1716 | _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 1717 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 1718 | (__v8sf) __W, |
| 1719 | (__mmask8) __U); |
| 1720 | } |
| 1721 | |
| 1722 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1723 | _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) { |
| 1724 | return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, |
| 1725 | (__v8sf) |
| 1726 | _mm256_setzero_ps (), |
| 1727 | (__mmask8) __U); |
| 1728 | } |
| 1729 | |
| 1730 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1731 | _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 1732 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 1733 | (__v4si) __W, |
| 1734 | (__mmask8) __U); |
| 1735 | } |
| 1736 | |
| 1737 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1738 | _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) { |
| 1739 | return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, |
| 1740 | (__v4si) |
| 1741 | _mm_setzero_si128 (), |
| 1742 | (__mmask8) __U); |
| 1743 | } |
| 1744 | |
| 1745 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1746 | _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 1747 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 1748 | (__v8si) __W, |
| 1749 | (__mmask8) __U); |
| 1750 | } |
| 1751 | |
| 1752 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1753 | _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) { |
| 1754 | return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, |
| 1755 | (__v8si) |
| 1756 | _mm256_setzero_si256 (), |
| 1757 | (__mmask8) __U); |
| 1758 | } |
| 1759 | |
| 1760 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1761 | _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) { |
| 1762 | __builtin_ia32_compressstoredf128_mask ((__v2df *) __P, |
| 1763 | (__v2df) __A, |
| 1764 | (__mmask8) __U); |
| 1765 | } |
| 1766 | |
| 1767 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1768 | _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) { |
| 1769 | __builtin_ia32_compressstoredf256_mask ((__v4df *) __P, |
| 1770 | (__v4df) __A, |
| 1771 | (__mmask8) __U); |
| 1772 | } |
| 1773 | |
| 1774 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1775 | _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) { |
| 1776 | __builtin_ia32_compressstoredi128_mask ((__v2di *) __P, |
| 1777 | (__v2di) __A, |
| 1778 | (__mmask8) __U); |
| 1779 | } |
| 1780 | |
| 1781 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1782 | _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) { |
| 1783 | __builtin_ia32_compressstoredi256_mask ((__v4di *) __P, |
| 1784 | (__v4di) __A, |
| 1785 | (__mmask8) __U); |
| 1786 | } |
| 1787 | |
| 1788 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1789 | _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) { |
| 1790 | __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, |
| 1791 | (__v4sf) __A, |
| 1792 | (__mmask8) __U); |
| 1793 | } |
| 1794 | |
| 1795 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1796 | _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) { |
| 1797 | __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P, |
| 1798 | (__v8sf) __A, |
| 1799 | (__mmask8) __U); |
| 1800 | } |
| 1801 | |
| 1802 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1803 | _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) { |
| 1804 | __builtin_ia32_compressstoresi128_mask ((__v4si *) __P, |
| 1805 | (__v4si) __A, |
| 1806 | (__mmask8) __U); |
| 1807 | } |
| 1808 | |
| 1809 | static __inline__ void __DEFAULT_FN_ATTRS |
| 1810 | _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) { |
| 1811 | __builtin_ia32_compressstoresi256_mask ((__v8si *) __P, |
| 1812 | (__v8si) __A, |
| 1813 | (__mmask8) __U); |
| 1814 | } |
| 1815 | |
| 1816 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1817 | _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 1818 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U, |
| 1819 | (__v2df)_mm_cvtepi32_pd(__A), |
| 1820 | (__v2df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1821 | } |
| 1822 | |
| 1823 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 1824 | _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 1825 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U, |
| 1826 | (__v2df)_mm_cvtepi32_pd(__A), |
| 1827 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1828 | } |
| 1829 | |
| 1830 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1831 | _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 1832 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U, |
| 1833 | (__v4df)_mm256_cvtepi32_pd(__A), |
| 1834 | (__v4df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1835 | } |
| 1836 | |
| 1837 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 1838 | _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 1839 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U, |
| 1840 | (__v4df)_mm256_cvtepi32_pd(__A), |
| 1841 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1842 | } |
| 1843 | |
| 1844 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1845 | _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1846 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1847 | (__v4sf)_mm_cvtepi32_ps(__A), |
| 1848 | (__v4sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1849 | } |
| 1850 | |
| 1851 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1852 | _mm_maskz_cvtepi32_ps (__mmask16 __U, __m128i __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1853 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1854 | (__v4sf)_mm_cvtepi32_ps(__A), |
| 1855 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1856 | } |
| 1857 | |
| 1858 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1859 | _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1860 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1861 | (__v8sf)_mm256_cvtepi32_ps(__A), |
| 1862 | (__v8sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1863 | } |
| 1864 | |
| 1865 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 1866 | _mm256_maskz_cvtepi32_ps (__mmask16 __U, __m256i __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1867 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 1868 | (__v8sf)_mm256_cvtepi32_ps(__A), |
| 1869 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1870 | } |
| 1871 | |
| 1872 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1873 | _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 1874 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 1875 | (__v4si) __W, |
| 1876 | (__mmask8) __U); |
| 1877 | } |
| 1878 | |
| 1879 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1880 | _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) { |
| 1881 | return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, |
| 1882 | (__v4si) |
| 1883 | _mm_setzero_si128 (), |
| 1884 | (__mmask8) __U); |
| 1885 | } |
| 1886 | |
| 1887 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1888 | _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1889 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 1890 | (__v4si)_mm256_cvtpd_epi32(__A), |
| 1891 | (__v4si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1892 | } |
| 1893 | |
| 1894 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1895 | _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1896 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 1897 | (__v4si)_mm256_cvtpd_epi32(__A), |
| 1898 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1899 | } |
| 1900 | |
| 1901 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1902 | _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) { |
| 1903 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 1904 | (__v4sf) __W, |
| 1905 | (__mmask8) __U); |
| 1906 | } |
| 1907 | |
| 1908 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1909 | _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) { |
| 1910 | return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, |
| 1911 | (__v4sf) |
| 1912 | _mm_setzero_ps (), |
| 1913 | (__mmask8) __U); |
| 1914 | } |
| 1915 | |
| 1916 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1917 | _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1918 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1919 | (__v4sf)_mm256_cvtpd_ps(__A), |
| 1920 | (__v4sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1921 | } |
| 1922 | |
| 1923 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 1924 | _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1925 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 1926 | (__v4sf)_mm256_cvtpd_ps(__A), |
| 1927 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1928 | } |
| 1929 | |
| 1930 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1931 | _mm_cvtpd_epu32 (__m128d __A) { |
| 1932 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 1933 | (__v4si) |
| 1934 | _mm_setzero_si128 (), |
| 1935 | (__mmask8) -1); |
| 1936 | } |
| 1937 | |
| 1938 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1939 | _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 1940 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 1941 | (__v4si) __W, |
| 1942 | (__mmask8) __U); |
| 1943 | } |
| 1944 | |
| 1945 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1946 | _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) { |
| 1947 | return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, |
| 1948 | (__v4si) |
| 1949 | _mm_setzero_si128 (), |
| 1950 | (__mmask8) __U); |
| 1951 | } |
| 1952 | |
| 1953 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1954 | _mm256_cvtpd_epu32 (__m256d __A) { |
| 1955 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 1956 | (__v4si) |
| 1957 | _mm_setzero_si128 (), |
| 1958 | (__mmask8) -1); |
| 1959 | } |
| 1960 | |
| 1961 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1962 | _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 1963 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 1964 | (__v4si) __W, |
| 1965 | (__mmask8) __U); |
| 1966 | } |
| 1967 | |
| 1968 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1969 | _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) { |
| 1970 | return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, |
| 1971 | (__v4si) |
| 1972 | _mm_setzero_si128 (), |
| 1973 | (__mmask8) __U); |
| 1974 | } |
| 1975 | |
| 1976 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1977 | _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1978 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 1979 | (__v4si)_mm_cvtps_epi32(__A), |
| 1980 | (__v4si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1981 | } |
| 1982 | |
| 1983 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 1984 | _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1985 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 1986 | (__v4si)_mm_cvtps_epi32(__A), |
| 1987 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1988 | } |
| 1989 | |
| 1990 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1991 | _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1992 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 1993 | (__v8si)_mm256_cvtps_epi32(__A), |
| 1994 | (__v8si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 1995 | } |
| 1996 | |
| 1997 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 1998 | _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 1999 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 2000 | (__v8si)_mm256_cvtps_epi32(__A), |
| 2001 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2002 | } |
| 2003 | |
| 2004 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2005 | _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2006 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2007 | (__v2df)_mm_cvtps_pd(__A), |
| 2008 | (__v2df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2009 | } |
| 2010 | |
| 2011 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2012 | _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2013 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2014 | (__v2df)_mm_cvtps_pd(__A), |
| 2015 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2016 | } |
| 2017 | |
| 2018 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2019 | _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2020 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2021 | (__v4df)_mm256_cvtps_pd(__A), |
| 2022 | (__v4df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2023 | } |
| 2024 | |
| 2025 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2026 | _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2027 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2028 | (__v4df)_mm256_cvtps_pd(__A), |
| 2029 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2030 | } |
| 2031 | |
| 2032 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2033 | _mm_cvtps_epu32 (__m128 __A) { |
| 2034 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2035 | (__v4si) |
| 2036 | _mm_setzero_si128 (), |
| 2037 | (__mmask8) -1); |
| 2038 | } |
| 2039 | |
| 2040 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2041 | _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2042 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2043 | (__v4si) __W, |
| 2044 | (__mmask8) __U); |
| 2045 | } |
| 2046 | |
| 2047 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2048 | _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) { |
| 2049 | return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, |
| 2050 | (__v4si) |
| 2051 | _mm_setzero_si128 (), |
| 2052 | (__mmask8) __U); |
| 2053 | } |
| 2054 | |
| 2055 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2056 | _mm256_cvtps_epu32 (__m256 __A) { |
| 2057 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2058 | (__v8si) |
| 2059 | _mm256_setzero_si256 (), |
| 2060 | (__mmask8) -1); |
| 2061 | } |
| 2062 | |
| 2063 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2064 | _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2065 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2066 | (__v8si) __W, |
| 2067 | (__mmask8) __U); |
| 2068 | } |
| 2069 | |
| 2070 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2071 | _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) { |
| 2072 | return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, |
| 2073 | (__v8si) |
| 2074 | _mm256_setzero_si256 (), |
| 2075 | (__mmask8) __U); |
| 2076 | } |
| 2077 | |
| 2078 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2079 | _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2080 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2081 | (__v4si) __W, |
| 2082 | (__mmask8) __U); |
| 2083 | } |
| 2084 | |
| 2085 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2086 | _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) { |
| 2087 | return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, |
| 2088 | (__v4si) |
| 2089 | _mm_setzero_si128 (), |
| 2090 | (__mmask8) __U); |
| 2091 | } |
| 2092 | |
| 2093 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2094 | _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2095 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2096 | (__v4si)_mm256_cvttpd_epi32(__A), |
| 2097 | (__v4si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2098 | } |
| 2099 | |
| 2100 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2101 | _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2102 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2103 | (__v4si)_mm256_cvttpd_epi32(__A), |
| 2104 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2105 | } |
| 2106 | |
| 2107 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2108 | _mm_cvttpd_epu32 (__m128d __A) { |
| 2109 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2110 | (__v4si) |
| 2111 | _mm_setzero_si128 (), |
| 2112 | (__mmask8) -1); |
| 2113 | } |
| 2114 | |
| 2115 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2116 | _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 2117 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2118 | (__v4si) __W, |
| 2119 | (__mmask8) __U); |
| 2120 | } |
| 2121 | |
| 2122 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2123 | _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) { |
| 2124 | return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, |
| 2125 | (__v4si) |
| 2126 | _mm_setzero_si128 (), |
| 2127 | (__mmask8) __U); |
| 2128 | } |
| 2129 | |
| 2130 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2131 | _mm256_cvttpd_epu32 (__m256d __A) { |
| 2132 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2133 | (__v4si) |
| 2134 | _mm_setzero_si128 (), |
| 2135 | (__mmask8) -1); |
| 2136 | } |
| 2137 | |
| 2138 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2139 | _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) { |
| 2140 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2141 | (__v4si) __W, |
| 2142 | (__mmask8) __U); |
| 2143 | } |
| 2144 | |
| 2145 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2146 | _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) { |
| 2147 | return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, |
| 2148 | (__v4si) |
| 2149 | _mm_setzero_si128 (), |
| 2150 | (__mmask8) __U); |
| 2151 | } |
| 2152 | |
| 2153 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2154 | _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2155 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2156 | (__v4si)_mm_cvttps_epi32(__A), |
| 2157 | (__v4si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2158 | } |
| 2159 | |
| 2160 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2161 | _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2162 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2163 | (__v4si)_mm_cvttps_epi32(__A), |
| 2164 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2165 | } |
| 2166 | |
| 2167 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2168 | _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2169 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 2170 | (__v8si)_mm256_cvttps_epi32(__A), |
| 2171 | (__v8si)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2172 | } |
| 2173 | |
| 2174 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2175 | _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) { |
Craig Topper | 21f66a3 | 2018-02-24 18:55:13 +0000 | [diff] [blame] | 2176 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 2177 | (__v8si)_mm256_cvttps_epi32(__A), |
| 2178 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2179 | } |
| 2180 | |
| 2181 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2182 | _mm_cvttps_epu32 (__m128 __A) { |
| 2183 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2184 | (__v4si) |
| 2185 | _mm_setzero_si128 (), |
| 2186 | (__mmask8) -1); |
| 2187 | } |
| 2188 | |
| 2189 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2190 | _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 2191 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2192 | (__v4si) __W, |
| 2193 | (__mmask8) __U); |
| 2194 | } |
| 2195 | |
| 2196 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2197 | _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) { |
| 2198 | return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, |
| 2199 | (__v4si) |
| 2200 | _mm_setzero_si128 (), |
| 2201 | (__mmask8) __U); |
| 2202 | } |
| 2203 | |
| 2204 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2205 | _mm256_cvttps_epu32 (__m256 __A) { |
| 2206 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2207 | (__v8si) |
| 2208 | _mm256_setzero_si256 (), |
| 2209 | (__mmask8) -1); |
| 2210 | } |
| 2211 | |
| 2212 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2213 | _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) { |
| 2214 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2215 | (__v8si) __W, |
| 2216 | (__mmask8) __U); |
| 2217 | } |
| 2218 | |
| 2219 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2220 | _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) { |
| 2221 | return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, |
| 2222 | (__v8si) |
| 2223 | _mm256_setzero_si256 (), |
| 2224 | (__mmask8) __U); |
| 2225 | } |
| 2226 | |
| 2227 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2228 | _mm_cvtepu32_pd (__m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2229 | return (__m128d) __builtin_convertvector( |
| 2230 | __builtin_shufflevector((__v4su)__A, (__v4su)__A, 0, 1), __v2df); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2231 | } |
| 2232 | |
| 2233 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2234 | _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2235 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U, |
| 2236 | (__v2df)_mm_cvtepu32_pd(__A), |
| 2237 | (__v2df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2238 | } |
| 2239 | |
| 2240 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2241 | _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2242 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U, |
| 2243 | (__v2df)_mm_cvtepu32_pd(__A), |
| 2244 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2245 | } |
| 2246 | |
| 2247 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2248 | _mm256_cvtepu32_pd (__m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2249 | return (__m256d)__builtin_convertvector((__v4su)__A, __v4df); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2250 | } |
| 2251 | |
| 2252 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2253 | _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2254 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U, |
| 2255 | (__v4df)_mm256_cvtepu32_pd(__A), |
| 2256 | (__v4df)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2257 | } |
| 2258 | |
| 2259 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2260 | _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) { |
Simon Pilgrim | 698528d | 2016-11-16 09:27:40 +0000 | [diff] [blame] | 2261 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U, |
| 2262 | (__v4df)_mm256_cvtepu32_pd(__A), |
| 2263 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2264 | } |
| 2265 | |
| 2266 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2267 | _mm_cvtepu32_ps (__m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2268 | return (__m128)__builtin_convertvector((__v4su)__A, __v4sf); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2269 | } |
| 2270 | |
| 2271 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2272 | _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2273 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2274 | (__v4sf)_mm_cvtepu32_ps(__A), |
| 2275 | (__v4sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2276 | } |
| 2277 | |
| 2278 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2279 | _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2280 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2281 | (__v4sf)_mm_cvtepu32_ps(__A), |
| 2282 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2283 | } |
| 2284 | |
| 2285 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2286 | _mm256_cvtepu32_ps (__m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2287 | return (__m256)__builtin_convertvector((__v8su)__A, __v8sf); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2288 | } |
| 2289 | |
| 2290 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2291 | _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2292 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2293 | (__v8sf)_mm256_cvtepu32_ps(__A), |
| 2294 | (__v8sf)__W); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2295 | } |
| 2296 | |
| 2297 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2298 | _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) { |
Craig Topper | 842171d | 2018-05-21 20:19:17 +0000 | [diff] [blame] | 2299 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2300 | (__v8sf)_mm256_cvtepu32_ps(__A), |
| 2301 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 74da387 | 2015-07-28 08:26:14 +0000 | [diff] [blame] | 2302 | } |
| 2303 | |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2304 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2305 | _mm_mask_div_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2306 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2307 | (__v2df)_mm_div_pd(__A, __B), |
| 2308 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2309 | } |
| 2310 | |
| 2311 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2312 | _mm_maskz_div_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 2313 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2314 | (__v2df)_mm_div_pd(__A, __B), |
| 2315 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2316 | } |
| 2317 | |
| 2318 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2319 | _mm256_mask_div_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 2320 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2321 | (__v4df)_mm256_div_pd(__A, __B), |
| 2322 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2323 | } |
| 2324 | |
| 2325 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2326 | _mm256_maskz_div_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 2327 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2328 | (__v4df)_mm256_div_pd(__A, __B), |
| 2329 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2330 | } |
| 2331 | |
| 2332 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2333 | _mm_mask_div_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2334 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2335 | (__v4sf)_mm_div_ps(__A, __B), |
| 2336 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2337 | } |
| 2338 | |
| 2339 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2340 | _mm_maskz_div_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 2341 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2342 | (__v4sf)_mm_div_ps(__A, __B), |
| 2343 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2344 | } |
| 2345 | |
| 2346 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2347 | _mm256_mask_div_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2348 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2349 | (__v8sf)_mm256_div_ps(__A, __B), |
| 2350 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2351 | } |
| 2352 | |
| 2353 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2354 | _mm256_maskz_div_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 2355 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2356 | (__v8sf)_mm256_div_ps(__A, __B), |
| 2357 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2358 | } |
| 2359 | |
| 2360 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2361 | _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2362 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2363 | (__v2df) __W, |
| 2364 | (__mmask8) __U); |
| 2365 | } |
| 2366 | |
| 2367 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2368 | _mm_maskz_expand_pd (__mmask8 __U, __m128d __A) { |
| 2369 | return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, |
| 2370 | (__v2df) |
| 2371 | _mm_setzero_pd (), |
| 2372 | (__mmask8) __U); |
| 2373 | } |
| 2374 | |
| 2375 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2376 | _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2377 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2378 | (__v4df) __W, |
| 2379 | (__mmask8) __U); |
| 2380 | } |
| 2381 | |
| 2382 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2383 | _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) { |
| 2384 | return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, |
| 2385 | (__v4df) |
| 2386 | _mm256_setzero_pd (), |
| 2387 | (__mmask8) __U); |
| 2388 | } |
| 2389 | |
| 2390 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2391 | _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2392 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2393 | (__v2di) __W, |
| 2394 | (__mmask8) __U); |
| 2395 | } |
| 2396 | |
| 2397 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2398 | _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) { |
| 2399 | return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, |
| 2400 | (__v2di) |
| 2401 | _mm_setzero_si128 (), |
| 2402 | (__mmask8) __U); |
| 2403 | } |
| 2404 | |
| 2405 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2406 | _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2407 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2408 | (__v4di) __W, |
| 2409 | (__mmask8) __U); |
| 2410 | } |
| 2411 | |
| 2412 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2413 | _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) { |
| 2414 | return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, |
| 2415 | (__v4di) |
| 2416 | _mm256_setzero_si256 (), |
| 2417 | (__mmask8) __U); |
| 2418 | } |
| 2419 | |
| 2420 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2421 | _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) { |
| 2422 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2423 | (__v2df) __W, |
| 2424 | (__mmask8) |
| 2425 | __U); |
| 2426 | } |
| 2427 | |
| 2428 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2429 | _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2430 | return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P, |
| 2431 | (__v2df) |
| 2432 | _mm_setzero_pd (), |
| 2433 | (__mmask8) |
| 2434 | __U); |
| 2435 | } |
| 2436 | |
| 2437 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2438 | _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) { |
| 2439 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2440 | (__v4df) __W, |
| 2441 | (__mmask8) |
| 2442 | __U); |
| 2443 | } |
| 2444 | |
| 2445 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2446 | _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) { |
| 2447 | return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P, |
| 2448 | (__v4df) |
| 2449 | _mm256_setzero_pd (), |
| 2450 | (__mmask8) |
| 2451 | __U); |
| 2452 | } |
| 2453 | |
| 2454 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2455 | _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2456 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2457 | (__v2di) __W, |
| 2458 | (__mmask8) |
| 2459 | __U); |
| 2460 | } |
| 2461 | |
| 2462 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2463 | _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2464 | return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P, |
| 2465 | (__v2di) |
| 2466 | _mm_setzero_si128 (), |
| 2467 | (__mmask8) |
| 2468 | __U); |
| 2469 | } |
| 2470 | |
| 2471 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2472 | _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U, |
| 2473 | void const *__P) { |
| 2474 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2475 | (__v4di) __W, |
| 2476 | (__mmask8) |
| 2477 | __U); |
| 2478 | } |
| 2479 | |
| 2480 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2481 | _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) { |
| 2482 | return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P, |
| 2483 | (__v4di) |
| 2484 | _mm256_setzero_si256 (), |
| 2485 | (__mmask8) |
| 2486 | __U); |
| 2487 | } |
| 2488 | |
| 2489 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2490 | _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) { |
| 2491 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2492 | (__v4sf) __W, |
| 2493 | (__mmask8) __U); |
| 2494 | } |
| 2495 | |
| 2496 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2497 | _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2498 | return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, |
| 2499 | (__v4sf) |
| 2500 | _mm_setzero_ps (), |
| 2501 | (__mmask8) |
| 2502 | __U); |
| 2503 | } |
| 2504 | |
| 2505 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2506 | _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) { |
| 2507 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2508 | (__v8sf) __W, |
| 2509 | (__mmask8) __U); |
| 2510 | } |
| 2511 | |
| 2512 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2513 | _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) { |
| 2514 | return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P, |
| 2515 | (__v8sf) |
| 2516 | _mm256_setzero_ps (), |
| 2517 | (__mmask8) |
| 2518 | __U); |
| 2519 | } |
| 2520 | |
| 2521 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2522 | _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) { |
| 2523 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2524 | (__v4si) __W, |
| 2525 | (__mmask8) |
| 2526 | __U); |
| 2527 | } |
| 2528 | |
| 2529 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2530 | _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2531 | return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P, |
| 2532 | (__v4si) |
| 2533 | _mm_setzero_si128 (), |
| 2534 | (__mmask8) __U); |
| 2535 | } |
| 2536 | |
| 2537 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2538 | _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U, |
| 2539 | void const *__P) { |
| 2540 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2541 | (__v8si) __W, |
| 2542 | (__mmask8) |
| 2543 | __U); |
| 2544 | } |
| 2545 | |
| 2546 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2547 | _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) { |
| 2548 | return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P, |
| 2549 | (__v8si) |
| 2550 | _mm256_setzero_si256 (), |
| 2551 | (__mmask8) |
| 2552 | __U); |
| 2553 | } |
| 2554 | |
| 2555 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2556 | _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2557 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2558 | (__v4sf) __W, |
| 2559 | (__mmask8) __U); |
| 2560 | } |
| 2561 | |
| 2562 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2563 | _mm_maskz_expand_ps (__mmask8 __U, __m128 __A) { |
| 2564 | return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, |
| 2565 | (__v4sf) |
| 2566 | _mm_setzero_ps (), |
| 2567 | (__mmask8) __U); |
| 2568 | } |
| 2569 | |
| 2570 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2571 | _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2572 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 2573 | (__v8sf) __W, |
| 2574 | (__mmask8) __U); |
| 2575 | } |
| 2576 | |
| 2577 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2578 | _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) { |
| 2579 | return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, |
| 2580 | (__v8sf) |
| 2581 | _mm256_setzero_ps (), |
| 2582 | (__mmask8) __U); |
| 2583 | } |
| 2584 | |
| 2585 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2586 | _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) { |
| 2587 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 2588 | (__v4si) __W, |
| 2589 | (__mmask8) __U); |
| 2590 | } |
| 2591 | |
| 2592 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2593 | _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) { |
| 2594 | return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, |
| 2595 | (__v4si) |
| 2596 | _mm_setzero_si128 (), |
| 2597 | (__mmask8) __U); |
| 2598 | } |
| 2599 | |
| 2600 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2601 | _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) { |
| 2602 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 2603 | (__v8si) __W, |
| 2604 | (__mmask8) __U); |
| 2605 | } |
| 2606 | |
| 2607 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2608 | _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) { |
| 2609 | return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, |
| 2610 | (__v8si) |
| 2611 | _mm256_setzero_si256 (), |
| 2612 | (__mmask8) __U); |
| 2613 | } |
| 2614 | |
| 2615 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2616 | _mm_getexp_pd (__m128d __A) { |
| 2617 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2618 | (__v2df) |
| 2619 | _mm_setzero_pd (), |
| 2620 | (__mmask8) -1); |
| 2621 | } |
| 2622 | |
| 2623 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2624 | _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) { |
| 2625 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2626 | (__v2df) __W, |
| 2627 | (__mmask8) __U); |
| 2628 | } |
| 2629 | |
| 2630 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 2631 | _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) { |
| 2632 | return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, |
| 2633 | (__v2df) |
| 2634 | _mm_setzero_pd (), |
| 2635 | (__mmask8) __U); |
| 2636 | } |
| 2637 | |
| 2638 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2639 | _mm256_getexp_pd (__m256d __A) { |
| 2640 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 2641 | (__v4df) |
| 2642 | _mm256_setzero_pd (), |
| 2643 | (__mmask8) -1); |
| 2644 | } |
| 2645 | |
| 2646 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2647 | _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) { |
| 2648 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 2649 | (__v4df) __W, |
| 2650 | (__mmask8) __U); |
| 2651 | } |
| 2652 | |
| 2653 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 2654 | _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) { |
| 2655 | return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, |
| 2656 | (__v4df) |
| 2657 | _mm256_setzero_pd (), |
| 2658 | (__mmask8) __U); |
| 2659 | } |
| 2660 | |
| 2661 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2662 | _mm_getexp_ps (__m128 __A) { |
| 2663 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 2664 | (__v4sf) |
| 2665 | _mm_setzero_ps (), |
| 2666 | (__mmask8) -1); |
| 2667 | } |
| 2668 | |
| 2669 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2670 | _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) { |
| 2671 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 2672 | (__v4sf) __W, |
| 2673 | (__mmask8) __U); |
| 2674 | } |
| 2675 | |
| 2676 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 2677 | _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) { |
| 2678 | return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, |
| 2679 | (__v4sf) |
| 2680 | _mm_setzero_ps (), |
| 2681 | (__mmask8) __U); |
| 2682 | } |
| 2683 | |
| 2684 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2685 | _mm256_getexp_ps (__m256 __A) { |
| 2686 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 2687 | (__v8sf) |
| 2688 | _mm256_setzero_ps (), |
| 2689 | (__mmask8) -1); |
| 2690 | } |
| 2691 | |
| 2692 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2693 | _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) { |
| 2694 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 2695 | (__v8sf) __W, |
| 2696 | (__mmask8) __U); |
| 2697 | } |
| 2698 | |
| 2699 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 2700 | _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) { |
| 2701 | return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, |
| 2702 | (__v8sf) |
| 2703 | _mm256_setzero_ps (), |
| 2704 | (__mmask8) __U); |
| 2705 | } |
| 2706 | |
| 2707 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2708 | _mm_mask_max_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2709 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2710 | (__v2df)_mm_max_pd(__A, __B), |
| 2711 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2712 | } |
| 2713 | |
| 2714 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2715 | _mm_maskz_max_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 2716 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2717 | (__v2df)_mm_max_pd(__A, __B), |
| 2718 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2719 | } |
| 2720 | |
| 2721 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2722 | _mm256_mask_max_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 2723 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2724 | (__v4df)_mm256_max_pd(__A, __B), |
| 2725 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2726 | } |
| 2727 | |
| 2728 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2729 | _mm256_maskz_max_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 2730 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2731 | (__v4df)_mm256_max_pd(__A, __B), |
| 2732 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2733 | } |
| 2734 | |
| 2735 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2736 | _mm_mask_max_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2737 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2738 | (__v4sf)_mm_max_ps(__A, __B), |
| 2739 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2740 | } |
| 2741 | |
| 2742 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2743 | _mm_maskz_max_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 2744 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2745 | (__v4sf)_mm_max_ps(__A, __B), |
| 2746 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2747 | } |
| 2748 | |
| 2749 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2750 | _mm256_mask_max_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2751 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2752 | (__v8sf)_mm256_max_ps(__A, __B), |
| 2753 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2754 | } |
| 2755 | |
| 2756 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2757 | _mm256_maskz_max_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 2758 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2759 | (__v8sf)_mm256_max_ps(__A, __B), |
| 2760 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2761 | } |
| 2762 | |
| 2763 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2764 | _mm_mask_min_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2765 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2766 | (__v2df)_mm_min_pd(__A, __B), |
| 2767 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2768 | } |
| 2769 | |
| 2770 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2771 | _mm_maskz_min_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 2772 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2773 | (__v2df)_mm_min_pd(__A, __B), |
| 2774 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2775 | } |
| 2776 | |
| 2777 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2778 | _mm256_mask_min_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 2779 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2780 | (__v4df)_mm256_min_pd(__A, __B), |
| 2781 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2782 | } |
| 2783 | |
| 2784 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2785 | _mm256_maskz_min_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 2786 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2787 | (__v4df)_mm256_min_pd(__A, __B), |
| 2788 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2789 | } |
| 2790 | |
| 2791 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2792 | _mm_mask_min_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2793 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2794 | (__v4sf)_mm_min_ps(__A, __B), |
| 2795 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2796 | } |
| 2797 | |
| 2798 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2799 | _mm_maskz_min_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 2800 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2801 | (__v4sf)_mm_min_ps(__A, __B), |
| 2802 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2803 | } |
| 2804 | |
| 2805 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2806 | _mm256_mask_min_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2807 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2808 | (__v8sf)_mm256_min_ps(__A, __B), |
| 2809 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2810 | } |
| 2811 | |
| 2812 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 160ca84 | 2016-10-29 19:02:03 +0000 | [diff] [blame] | 2813 | _mm256_maskz_min_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 2814 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2815 | (__v8sf)_mm256_min_ps(__A, __B), |
| 2816 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2817 | } |
| 2818 | |
| 2819 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2820 | _mm_mask_mul_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 2821 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2822 | (__v2df)_mm_mul_pd(__A, __B), |
| 2823 | (__v2df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2824 | } |
| 2825 | |
| 2826 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2827 | _mm_maskz_mul_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 2828 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 2829 | (__v2df)_mm_mul_pd(__A, __B), |
| 2830 | (__v2df)_mm_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2831 | } |
| 2832 | |
| 2833 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2834 | _mm256_mask_mul_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 2835 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2836 | (__v4df)_mm256_mul_pd(__A, __B), |
| 2837 | (__v4df)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2838 | } |
| 2839 | |
| 2840 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2841 | _mm256_maskz_mul_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 2842 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 2843 | (__v4df)_mm256_mul_pd(__A, __B), |
| 2844 | (__v4df)_mm256_setzero_pd()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2845 | } |
| 2846 | |
| 2847 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2848 | _mm_mask_mul_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 2849 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2850 | (__v4sf)_mm_mul_ps(__A, __B), |
| 2851 | (__v4sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2852 | } |
| 2853 | |
| 2854 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2855 | _mm_maskz_mul_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 2856 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 2857 | (__v4sf)_mm_mul_ps(__A, __B), |
| 2858 | (__v4sf)_mm_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2859 | } |
| 2860 | |
| 2861 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2862 | _mm256_mask_mul_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 2863 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2864 | (__v8sf)_mm256_mul_ps(__A, __B), |
| 2865 | (__v8sf)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2866 | } |
| 2867 | |
| 2868 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 2dfab63 | 2016-09-04 18:30:17 +0000 | [diff] [blame] | 2869 | _mm256_maskz_mul_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 2870 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 2871 | (__v8sf)_mm256_mul_ps(__A, __B), |
| 2872 | (__v8sf)_mm256_setzero_ps()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2873 | } |
| 2874 | |
| 2875 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 78a9c40 | 2016-10-22 21:24:38 +0000 | [diff] [blame] | 2876 | _mm_mask_abs_epi32(__m128i __W, __mmask8 __U, __m128i __A) { |
| 2877 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2878 | (__v4si)_mm_abs_epi32(__A), |
| 2879 | (__v4si)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2880 | } |
| 2881 | |
| 2882 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 78a9c40 | 2016-10-22 21:24:38 +0000 | [diff] [blame] | 2883 | _mm_maskz_abs_epi32(__mmask8 __U, __m128i __A) { |
| 2884 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 2885 | (__v4si)_mm_abs_epi32(__A), |
| 2886 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2887 | } |
| 2888 | |
| 2889 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 78a9c40 | 2016-10-22 21:24:38 +0000 | [diff] [blame] | 2890 | _mm256_mask_abs_epi32(__m256i __W, __mmask8 __U, __m256i __A) { |
| 2891 | return (__m256i)__builtin_ia32_selectd_256((__mmask16)__U, |
| 2892 | (__v8si)_mm256_abs_epi32(__A), |
| 2893 | (__v8si)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2894 | } |
| 2895 | |
| 2896 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 78a9c40 | 2016-10-22 21:24:38 +0000 | [diff] [blame] | 2897 | _mm256_maskz_abs_epi32(__mmask8 __U, __m256i __A) { |
| 2898 | return (__m256i)__builtin_ia32_selectd_256((__mmask16)__U, |
| 2899 | (__v8si)_mm256_abs_epi32(__A), |
| 2900 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2901 | } |
| 2902 | |
| 2903 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2904 | _mm_abs_epi64 (__m128i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2905 | return (__m128i)__builtin_ia32_pabsq128((__v2di)__A); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2906 | } |
| 2907 | |
| 2908 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2909 | _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2910 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 2911 | (__v2di)_mm_abs_epi64(__A), |
| 2912 | (__v2di)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2913 | } |
| 2914 | |
| 2915 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2916 | _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2917 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 2918 | (__v2di)_mm_abs_epi64(__A), |
| 2919 | (__v2di)_mm_setzero_si128()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2920 | } |
| 2921 | |
| 2922 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2923 | _mm256_abs_epi64 (__m256i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2924 | return (__m256i)__builtin_ia32_pabsq256 ((__v4di)__A); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2925 | } |
| 2926 | |
| 2927 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2928 | _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2929 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 2930 | (__v4di)_mm256_abs_epi64(__A), |
| 2931 | (__v4di)__W); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2932 | } |
| 2933 | |
| 2934 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2935 | _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2936 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 2937 | (__v4di)_mm256_abs_epi64(__A), |
| 2938 | (__v4di)_mm256_setzero_si256()); |
Asaf Badouh | 78ee5cc | 2015-07-28 10:30:56 +0000 | [diff] [blame] | 2939 | } |
| 2940 | |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2941 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 2942 | _mm_maskz_max_epi32(__mmask8 __M, __m128i __A, __m128i __B) { |
| 2943 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 2944 | (__v4si)_mm_max_epi32(__A, __B), |
| 2945 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2946 | } |
| 2947 | |
| 2948 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 2949 | _mm_mask_max_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 2950 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 2951 | (__v4si)_mm_max_epi32(__A, __B), |
| 2952 | (__v4si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2953 | } |
| 2954 | |
| 2955 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 2956 | _mm256_maskz_max_epi32(__mmask8 __M, __m256i __A, __m256i __B) { |
| 2957 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 2958 | (__v8si)_mm256_max_epi32(__A, __B), |
| 2959 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2960 | } |
| 2961 | |
| 2962 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 2963 | _mm256_mask_max_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 2964 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 2965 | (__v8si)_mm256_max_epi32(__A, __B), |
| 2966 | (__v8si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2967 | } |
| 2968 | |
| 2969 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2970 | _mm_max_epi64 (__m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2971 | return (__m128i)__builtin_ia32_pmaxsq128((__v2di)__A, (__v2di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2972 | } |
| 2973 | |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2974 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2975 | _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 2976 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 2977 | (__v2di)_mm_max_epi64(__A, __B), |
| 2978 | (__v2di)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2979 | } |
| 2980 | |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2981 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 2982 | _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 2983 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 2984 | (__v2di)_mm_max_epi64(__A, __B), |
| 2985 | (__v2di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 2986 | } |
| 2987 | |
| 2988 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2989 | _mm256_max_epi64 (__m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 2990 | return (__m256i)__builtin_ia32_pmaxsq256((__v4di)__A, (__v4di)__B); |
| 2991 | } |
| 2992 | |
| 2993 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 2994 | _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 2995 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 2996 | (__v4di)_mm256_max_epi64(__A, __B), |
| 2997 | (__v4di)_mm256_setzero_si256()); |
| 2998 | } |
| 2999 | |
| 3000 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3001 | _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3002 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3003 | (__v4di)_mm256_max_epi64(__A, __B), |
| 3004 | (__v4di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3005 | } |
| 3006 | |
| 3007 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3008 | _mm_maskz_max_epu32(__mmask8 __M, __m128i __A, __m128i __B) { |
| 3009 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3010 | (__v4si)_mm_max_epu32(__A, __B), |
| 3011 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3012 | } |
| 3013 | |
| 3014 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3015 | _mm_mask_max_epu32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3016 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3017 | (__v4si)_mm_max_epu32(__A, __B), |
| 3018 | (__v4si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3019 | } |
| 3020 | |
| 3021 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3022 | _mm256_maskz_max_epu32(__mmask8 __M, __m256i __A, __m256i __B) { |
| 3023 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3024 | (__v8si)_mm256_max_epu32(__A, __B), |
| 3025 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3026 | } |
| 3027 | |
| 3028 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3029 | _mm256_mask_max_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3030 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3031 | (__v8si)_mm256_max_epu32(__A, __B), |
| 3032 | (__v8si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3033 | } |
| 3034 | |
| 3035 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3036 | _mm_max_epu64 (__m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3037 | return (__m128i)__builtin_ia32_pmaxuq128((__v2di)__A, (__v2di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3038 | } |
| 3039 | |
| 3040 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3041 | _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
| 3042 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3043 | (__v2di)_mm_max_epu64(__A, __B), |
| 3044 | (__v2di)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3045 | } |
| 3046 | |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3047 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3048 | _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3049 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3050 | (__v2di)_mm_max_epu64(__A, __B), |
| 3051 | (__v2di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3052 | } |
| 3053 | |
| 3054 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3055 | _mm256_max_epu64 (__m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3056 | return (__m256i)__builtin_ia32_pmaxuq256((__v4di)__A, (__v4di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3057 | } |
| 3058 | |
| 3059 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3060 | _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
| 3061 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3062 | (__v4di)_mm256_max_epu64(__A, __B), |
| 3063 | (__v4di)_mm256_setzero_si256()); |
| 3064 | } |
| 3065 | |
| 3066 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3067 | _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3068 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3069 | (__v4di)_mm256_max_epu64(__A, __B), |
| 3070 | (__v4di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3071 | } |
| 3072 | |
| 3073 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3074 | _mm_maskz_min_epi32(__mmask8 __M, __m128i __A, __m128i __B) { |
| 3075 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3076 | (__v4si)_mm_min_epi32(__A, __B), |
| 3077 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3078 | } |
| 3079 | |
| 3080 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3081 | _mm_mask_min_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3082 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3083 | (__v4si)_mm_min_epi32(__A, __B), |
| 3084 | (__v4si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3085 | } |
| 3086 | |
| 3087 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3088 | _mm256_maskz_min_epi32(__mmask8 __M, __m256i __A, __m256i __B) { |
| 3089 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3090 | (__v8si)_mm256_min_epi32(__A, __B), |
| 3091 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3092 | } |
| 3093 | |
| 3094 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3095 | _mm256_mask_min_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3096 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3097 | (__v8si)_mm256_min_epi32(__A, __B), |
| 3098 | (__v8si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3099 | } |
| 3100 | |
| 3101 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3102 | _mm_min_epi64 (__m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3103 | return (__m128i)__builtin_ia32_pminsq128((__v2di)__A, (__v2di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3104 | } |
| 3105 | |
| 3106 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3107 | _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3108 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3109 | (__v2di)_mm_min_epi64(__A, __B), |
| 3110 | (__v2di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3111 | } |
| 3112 | |
| 3113 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3114 | _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3115 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3116 | (__v2di)_mm_min_epi64(__A, __B), |
| 3117 | (__v2di)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3118 | } |
| 3119 | |
| 3120 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3121 | _mm256_min_epi64 (__m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3122 | return (__m256i)__builtin_ia32_pminsq256((__v4di)__A, (__v4di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3123 | } |
| 3124 | |
| 3125 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3126 | _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3127 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3128 | (__v4di)_mm256_min_epi64(__A, __B), |
| 3129 | (__v4di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3130 | } |
| 3131 | |
| 3132 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3133 | _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3134 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3135 | (__v4di)_mm256_min_epi64(__A, __B), |
| 3136 | (__v4di)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3137 | } |
| 3138 | |
| 3139 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3140 | _mm_maskz_min_epu32(__mmask8 __M, __m128i __A, __m128i __B) { |
| 3141 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3142 | (__v4si)_mm_min_epu32(__A, __B), |
| 3143 | (__v4si)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3144 | } |
| 3145 | |
| 3146 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3147 | _mm_mask_min_epu32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3148 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 3149 | (__v4si)_mm_min_epu32(__A, __B), |
| 3150 | (__v4si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3151 | } |
| 3152 | |
| 3153 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3154 | _mm256_maskz_min_epu32(__mmask8 __M, __m256i __A, __m256i __B) { |
| 3155 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3156 | (__v8si)_mm256_min_epu32(__A, __B), |
| 3157 | (__v8si)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3158 | } |
| 3159 | |
| 3160 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | eee7c05 | 2016-10-23 23:57:30 +0000 | [diff] [blame] | 3161 | _mm256_mask_min_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3162 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 3163 | (__v8si)_mm256_min_epu32(__A, __B), |
| 3164 | (__v8si)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3165 | } |
| 3166 | |
| 3167 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3168 | _mm_min_epu64 (__m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3169 | return (__m128i)__builtin_ia32_pminuq128((__v2di)__A, (__v2di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3170 | } |
| 3171 | |
| 3172 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3173 | _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) { |
| 3174 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3175 | (__v2di)_mm_min_epu64(__A, __B), |
| 3176 | (__v2di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3177 | } |
| 3178 | |
| 3179 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3180 | _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3181 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M, |
| 3182 | (__v2di)_mm_min_epu64(__A, __B), |
| 3183 | (__v2di)_mm_setzero_si128()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3184 | } |
| 3185 | |
| 3186 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3187 | _mm256_min_epu64 (__m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3188 | return (__m256i)__builtin_ia32_pminuq256((__v4di)__A, (__v4di)__B); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3189 | } |
| 3190 | |
| 3191 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3192 | _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) { |
| 3193 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3194 | (__v4di)_mm256_min_epu64(__A, __B), |
| 3195 | (__v4di)__W); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3196 | } |
| 3197 | |
| 3198 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3199 | _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) { |
Craig Topper | f2043b0 | 2018-05-23 04:51:54 +0000 | [diff] [blame] | 3200 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 3201 | (__v4di)_mm256_min_epu64(__A, __B), |
| 3202 | (__v4di)_mm256_setzero_si256()); |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3203 | } |
| 3204 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3205 | #define _mm_roundscale_pd(A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3206 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3207 | (int)(imm), \ |
| 3208 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3209 | (__mmask8)-1) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3210 | |
| 3211 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3212 | #define _mm_mask_roundscale_pd(W, U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3213 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3214 | (int)(imm), \ |
| 3215 | (__v2df)(__m128d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3216 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3217 | |
| 3218 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3219 | #define _mm_maskz_roundscale_pd(U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3220 | (__m128d)__builtin_ia32_rndscalepd_128_mask((__v2df)(__m128d)(A), \ |
| 3221 | (int)(imm), \ |
| 3222 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3223 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3224 | |
| 3225 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3226 | #define _mm256_roundscale_pd(A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3227 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3228 | (int)(imm), \ |
| 3229 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3230 | (__mmask8)-1) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3231 | |
| 3232 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3233 | #define _mm256_mask_roundscale_pd(W, U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3234 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3235 | (int)(imm), \ |
| 3236 | (__v4df)(__m256d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3237 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3238 | |
| 3239 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3240 | #define _mm256_maskz_roundscale_pd(U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3241 | (__m256d)__builtin_ia32_rndscalepd_256_mask((__v4df)(__m256d)(A), \ |
| 3242 | (int)(imm), \ |
| 3243 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3244 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3245 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3246 | #define _mm_roundscale_ps(A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3247 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3248 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3249 | (__mmask8)-1) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3250 | |
| 3251 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3252 | #define _mm_mask_roundscale_ps(W, U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3253 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3254 | (__v4sf)(__m128)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3255 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3256 | |
| 3257 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3258 | #define _mm_maskz_roundscale_ps(U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3259 | (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 3260 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3261 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3262 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3263 | #define _mm256_roundscale_ps(A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3264 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3265 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3266 | (__mmask8)-1) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3267 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3268 | #define _mm256_mask_roundscale_ps(W, U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3269 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3270 | (__v8sf)(__m256)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3271 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3272 | |
| 3273 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3274 | #define _mm256_maskz_roundscale_ps(U, A, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3275 | (__m256)__builtin_ia32_rndscaleps_256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 3276 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3277 | (__mmask8)(U)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3278 | |
| 3279 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3280 | _mm_scalef_pd (__m128d __A, __m128d __B) { |
| 3281 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3282 | (__v2df) __B, |
| 3283 | (__v2df) |
| 3284 | _mm_setzero_pd (), |
| 3285 | (__mmask8) -1); |
| 3286 | } |
| 3287 | |
| 3288 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3289 | _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A, |
| 3290 | __m128d __B) { |
| 3291 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3292 | (__v2df) __B, |
| 3293 | (__v2df) __W, |
| 3294 | (__mmask8) __U); |
| 3295 | } |
| 3296 | |
| 3297 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3298 | _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 3299 | return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, |
| 3300 | (__v2df) __B, |
| 3301 | (__v2df) |
| 3302 | _mm_setzero_pd (), |
| 3303 | (__mmask8) __U); |
| 3304 | } |
| 3305 | |
| 3306 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3307 | _mm256_scalef_pd (__m256d __A, __m256d __B) { |
| 3308 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3309 | (__v4df) __B, |
| 3310 | (__v4df) |
| 3311 | _mm256_setzero_pd (), |
| 3312 | (__mmask8) -1); |
| 3313 | } |
| 3314 | |
| 3315 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3316 | _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A, |
| 3317 | __m256d __B) { |
| 3318 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3319 | (__v4df) __B, |
| 3320 | (__v4df) __W, |
| 3321 | (__mmask8) __U); |
| 3322 | } |
| 3323 | |
| 3324 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3325 | _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) { |
| 3326 | return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, |
| 3327 | (__v4df) __B, |
| 3328 | (__v4df) |
| 3329 | _mm256_setzero_pd (), |
| 3330 | (__mmask8) __U); |
| 3331 | } |
| 3332 | |
| 3333 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3334 | _mm_scalef_ps (__m128 __A, __m128 __B) { |
| 3335 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3336 | (__v4sf) __B, |
| 3337 | (__v4sf) |
| 3338 | _mm_setzero_ps (), |
| 3339 | (__mmask8) -1); |
| 3340 | } |
| 3341 | |
| 3342 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3343 | _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3344 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3345 | (__v4sf) __B, |
| 3346 | (__v4sf) __W, |
| 3347 | (__mmask8) __U); |
| 3348 | } |
| 3349 | |
| 3350 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3351 | _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) { |
| 3352 | return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, |
| 3353 | (__v4sf) __B, |
| 3354 | (__v4sf) |
| 3355 | _mm_setzero_ps (), |
| 3356 | (__mmask8) __U); |
| 3357 | } |
| 3358 | |
| 3359 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3360 | _mm256_scalef_ps (__m256 __A, __m256 __B) { |
| 3361 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3362 | (__v8sf) __B, |
| 3363 | (__v8sf) |
| 3364 | _mm256_setzero_ps (), |
| 3365 | (__mmask8) -1); |
| 3366 | } |
| 3367 | |
| 3368 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3369 | _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A, |
| 3370 | __m256 __B) { |
| 3371 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3372 | (__v8sf) __B, |
| 3373 | (__v8sf) __W, |
| 3374 | (__mmask8) __U); |
| 3375 | } |
| 3376 | |
| 3377 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3378 | _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) { |
| 3379 | return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, |
| 3380 | (__v8sf) __B, |
| 3381 | (__v8sf) |
| 3382 | _mm256_setzero_ps (), |
| 3383 | (__mmask8) __U); |
| 3384 | } |
| 3385 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3386 | #define _mm_i64scatter_pd(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3387 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)-1, \ |
| 3388 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3389 | (__v2df)(__m128d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3390 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3391 | #define _mm_mask_i64scatter_pd(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3392 | __builtin_ia32_scatterdiv2df((double *)(addr), (__mmask8)(mask), \ |
| 3393 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3394 | (__v2df)(__m128d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3395 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3396 | #define _mm_i64scatter_epi64(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3397 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)-1, \ |
| 3398 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3399 | (__v2di)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3400 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3401 | #define _mm_mask_i64scatter_epi64(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3402 | __builtin_ia32_scatterdiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 3403 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3404 | (__v2di)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3405 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3406 | #define _mm256_i64scatter_pd(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3407 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)-1, \ |
| 3408 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3409 | (__v4df)(__m256d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3410 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3411 | #define _mm256_mask_i64scatter_pd(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3412 | __builtin_ia32_scatterdiv4df((double *)(addr), (__mmask8)(mask), \ |
| 3413 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3414 | (__v4df)(__m256d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3415 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3416 | #define _mm256_i64scatter_epi64(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3417 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)-1, \ |
| 3418 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3419 | (__v4di)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3420 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3421 | #define _mm256_mask_i64scatter_epi64(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3422 | __builtin_ia32_scatterdiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 3423 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3424 | (__v4di)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3425 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3426 | #define _mm_i64scatter_ps(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3427 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)-1, \ |
| 3428 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3429 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3430 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3431 | #define _mm_mask_i64scatter_ps(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3432 | __builtin_ia32_scatterdiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 3433 | (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3434 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3435 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3436 | #define _mm_i64scatter_epi32(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3437 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)-1, \ |
| 3438 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3439 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3440 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3441 | #define _mm_mask_i64scatter_epi32(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3442 | __builtin_ia32_scatterdiv4si((int *)(addr), (__mmask8)(mask), \ |
| 3443 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3444 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3445 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3446 | #define _mm256_i64scatter_ps(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3447 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)-1, \ |
| 3448 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3449 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3450 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3451 | #define _mm256_mask_i64scatter_ps(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3452 | __builtin_ia32_scatterdiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 3453 | (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3454 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3455 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3456 | #define _mm256_i64scatter_epi32(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3457 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)-1, \ |
| 3458 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3459 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3460 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3461 | #define _mm256_mask_i64scatter_epi32(addr, mask, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3462 | __builtin_ia32_scatterdiv8si((int *)(addr), (__mmask8)(mask), \ |
| 3463 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3464 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3465 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3466 | #define _mm_i32scatter_pd(addr, index, v1, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 3467 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)-1, \ |
| 3468 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3469 | (__v2df)(__m128d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3470 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3471 | #define _mm_mask_i32scatter_pd(addr, mask, index, v1, scale) \ |
| 3472 | __builtin_ia32_scattersiv2df((double *)(addr), (__mmask8)(mask), \ |
| 3473 | (__v4si)(__m128i)(index), \ |
| 3474 | (__v2df)(__m128d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3475 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3476 | #define _mm_i32scatter_epi64(addr, index, v1, scale) \ |
| 3477 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)-1, \ |
| 3478 | (__v4si)(__m128i)(index), \ |
| 3479 | (__v2di)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3480 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3481 | #define _mm_mask_i32scatter_epi64(addr, mask, index, v1, scale) \ |
| 3482 | __builtin_ia32_scattersiv2di((long long *)(addr), (__mmask8)(mask), \ |
| 3483 | (__v4si)(__m128i)(index), \ |
| 3484 | (__v2di)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3485 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3486 | #define _mm256_i32scatter_pd(addr, index, v1, scale) \ |
| 3487 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)-1, \ |
| 3488 | (__v4si)(__m128i)(index), \ |
| 3489 | (__v4df)(__m256d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3490 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3491 | #define _mm256_mask_i32scatter_pd(addr, mask, index, v1, scale) \ |
| 3492 | __builtin_ia32_scattersiv4df((double *)(addr), (__mmask8)(mask), \ |
| 3493 | (__v4si)(__m128i)(index), \ |
| 3494 | (__v4df)(__m256d)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3495 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3496 | #define _mm256_i32scatter_epi64(addr, index, v1, scale) \ |
| 3497 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)-1, \ |
| 3498 | (__v4si)(__m128i)(index), \ |
| 3499 | (__v4di)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3500 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3501 | #define _mm256_mask_i32scatter_epi64(addr, mask, index, v1, scale) \ |
| 3502 | __builtin_ia32_scattersiv4di((long long *)(addr), (__mmask8)(mask), \ |
| 3503 | (__v4si)(__m128i)(index), \ |
| 3504 | (__v4di)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3505 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3506 | #define _mm_i32scatter_ps(addr, index, v1, scale) \ |
| 3507 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)-1, \ |
| 3508 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3509 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3510 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3511 | #define _mm_mask_i32scatter_ps(addr, mask, index, v1, scale) \ |
| 3512 | __builtin_ia32_scattersiv4sf((float *)(addr), (__mmask8)(mask), \ |
| 3513 | (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \ |
| 3514 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3515 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3516 | #define _mm_i32scatter_epi32(addr, index, v1, scale) \ |
| 3517 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)-1, \ |
| 3518 | (__v4si)(__m128i)(index), \ |
| 3519 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3520 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3521 | #define _mm_mask_i32scatter_epi32(addr, mask, index, v1, scale) \ |
| 3522 | __builtin_ia32_scattersiv4si((int *)(addr), (__mmask8)(mask), \ |
| 3523 | (__v4si)(__m128i)(index), \ |
| 3524 | (__v4si)(__m128i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3525 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3526 | #define _mm256_i32scatter_ps(addr, index, v1, scale) \ |
| 3527 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)-1, \ |
| 3528 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 3529 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3530 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3531 | #define _mm256_mask_i32scatter_ps(addr, mask, index, v1, scale) \ |
| 3532 | __builtin_ia32_scattersiv8sf((float *)(addr), (__mmask8)(mask), \ |
| 3533 | (__v8si)(__m256i)(index), (__v8sf)(__m256)(v1), \ |
| 3534 | (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3535 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3536 | #define _mm256_i32scatter_epi32(addr, index, v1, scale) \ |
| 3537 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)-1, \ |
| 3538 | (__v8si)(__m256i)(index), \ |
| 3539 | (__v8si)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3540 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3541 | #define _mm256_mask_i32scatter_epi32(addr, mask, index, v1, scale) \ |
| 3542 | __builtin_ia32_scattersiv8si((int *)(addr), (__mmask8)(mask), \ |
| 3543 | (__v8si)(__m256i)(index), \ |
| 3544 | (__v8si)(__m256i)(v1), (int)(scale)) |
Asaf Badouh | b7cf71b | 2015-07-28 11:14:09 +0000 | [diff] [blame] | 3545 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3546 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3547 | _mm_mask_sqrt_pd(__m128d __W, __mmask8 __U, __m128d __A) { |
| 3548 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3549 | (__v2df)_mm_sqrt_pd(__A), |
| 3550 | (__v2df)__W); |
| 3551 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3552 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3553 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3554 | _mm_maskz_sqrt_pd(__mmask8 __U, __m128d __A) { |
| 3555 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3556 | (__v2df)_mm_sqrt_pd(__A), |
| 3557 | (__v2df)_mm_setzero_pd()); |
| 3558 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3559 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3560 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3561 | _mm256_mask_sqrt_pd(__m256d __W, __mmask8 __U, __m256d __A) { |
| 3562 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3563 | (__v4df)_mm256_sqrt_pd(__A), |
| 3564 | (__v4df)__W); |
| 3565 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3566 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3567 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3568 | _mm256_maskz_sqrt_pd(__mmask8 __U, __m256d __A) { |
| 3569 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3570 | (__v4df)_mm256_sqrt_pd(__A), |
| 3571 | (__v4df)_mm256_setzero_pd()); |
| 3572 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3573 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3574 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3575 | _mm_mask_sqrt_ps(__m128 __W, __mmask8 __U, __m128 __A) { |
| 3576 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3577 | (__v4sf)_mm_sqrt_ps(__A), |
| 3578 | (__v4sf)__W); |
| 3579 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3580 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3581 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3582 | _mm_maskz_sqrt_ps(__mmask8 __U, __m128 __A) { |
| 3583 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3584 | (__v4sf)_mm_sqrt_ps(__A), |
| 3585 | (__v4sf)_mm_setzero_pd()); |
| 3586 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3587 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3588 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3589 | _mm256_mask_sqrt_ps(__m256 __W, __mmask8 __U, __m256 __A) { |
| 3590 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3591 | (__v8sf)_mm256_sqrt_ps(__A), |
| 3592 | (__v8sf)__W); |
| 3593 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3594 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3595 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3596 | _mm256_maskz_sqrt_ps(__mmask8 __U, __m256 __A) { |
| 3597 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3598 | (__v8sf)_mm256_sqrt_ps(__A), |
| 3599 | (__v8sf)_mm256_setzero_ps()); |
| 3600 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3601 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3602 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3603 | _mm_mask_sub_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 3604 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3605 | (__v2df)_mm_sub_pd(__A, __B), |
| 3606 | (__v2df)__W); |
| 3607 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3608 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3609 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3610 | _mm_maskz_sub_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 3611 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 3612 | (__v2df)_mm_sub_pd(__A, __B), |
| 3613 | (__v2df)_mm_setzero_pd()); |
| 3614 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3615 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3616 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3617 | _mm256_mask_sub_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 3618 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3619 | (__v4df)_mm256_sub_pd(__A, __B), |
| 3620 | (__v4df)__W); |
| 3621 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3622 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3623 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3624 | _mm256_maskz_sub_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 3625 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 3626 | (__v4df)_mm256_sub_pd(__A, __B), |
| 3627 | (__v4df)_mm256_setzero_pd()); |
| 3628 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3629 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3630 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3631 | _mm_mask_sub_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 3632 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3633 | (__v4sf)_mm_sub_ps(__A, __B), |
| 3634 | (__v4sf)__W); |
| 3635 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3636 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3637 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3638 | _mm_maskz_sub_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 3639 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 3640 | (__v4sf)_mm_sub_ps(__A, __B), |
| 3641 | (__v4sf)_mm_setzero_ps()); |
| 3642 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3643 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3644 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3645 | _mm256_mask_sub_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 3646 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3647 | (__v8sf)_mm256_sub_ps(__A, __B), |
| 3648 | (__v8sf)__W); |
| 3649 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3650 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3651 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3652 | _mm256_maskz_sub_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 3653 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 3654 | (__v8sf)_mm256_sub_ps(__A, __B), |
| 3655 | (__v8sf)_mm256_setzero_ps()); |
| 3656 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3657 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3658 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3659 | _mm_permutex2var_epi32(__m128i __A, __m128i __I, __m128i __B) { |
| 3660 | return (__m128i)__builtin_ia32_vpermi2vard128((__v4si) __A, (__v4si)__I, |
| 3661 | (__v4si)__B); |
| 3662 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3663 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3664 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3665 | _mm_mask_permutex2var_epi32(__m128i __A, __mmask8 __U, __m128i __I, |
| 3666 | __m128i __B) { |
| 3667 | return (__m128i)__builtin_ia32_selectd_128(__U, |
| 3668 | (__v4si)_mm_permutex2var_epi32(__A, __I, __B), |
| 3669 | (__v4si)__A); |
| 3670 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3671 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3672 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3673 | _mm_mask2_permutex2var_epi32(__m128i __A, __m128i __I, __mmask8 __U, |
| 3674 | __m128i __B) { |
| 3675 | return (__m128i)__builtin_ia32_selectd_128(__U, |
| 3676 | (__v4si)_mm_permutex2var_epi32(__A, __I, __B), |
| 3677 | (__v4si)__I); |
| 3678 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3679 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3680 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3681 | _mm_maskz_permutex2var_epi32(__mmask8 __U, __m128i __A, __m128i __I, |
| 3682 | __m128i __B) { |
| 3683 | return (__m128i)__builtin_ia32_selectd_128(__U, |
| 3684 | (__v4si)_mm_permutex2var_epi32(__A, __I, __B), |
| 3685 | (__v4si)_mm_setzero_si128()); |
| 3686 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3687 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3688 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3689 | _mm256_permutex2var_epi32(__m256i __A, __m256i __I, __m256i __B) { |
| 3690 | return (__m256i)__builtin_ia32_vpermi2vard256((__v8si)__A, (__v8si) __I, |
| 3691 | (__v8si) __B); |
| 3692 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3693 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3694 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3695 | _mm256_mask_permutex2var_epi32(__m256i __A, __mmask8 __U, __m256i __I, |
| 3696 | __m256i __B) { |
| 3697 | return (__m256i)__builtin_ia32_selectd_256(__U, |
| 3698 | (__v8si)_mm256_permutex2var_epi32(__A, __I, __B), |
| 3699 | (__v8si)__A); |
| 3700 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3701 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3702 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3703 | _mm256_mask2_permutex2var_epi32(__m256i __A, __m256i __I, __mmask8 __U, |
| 3704 | __m256i __B) { |
| 3705 | return (__m256i)__builtin_ia32_selectd_256(__U, |
| 3706 | (__v8si)_mm256_permutex2var_epi32(__A, __I, __B), |
| 3707 | (__v8si)__I); |
| 3708 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3709 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3710 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3711 | _mm256_maskz_permutex2var_epi32(__mmask8 __U, __m256i __A, __m256i __I, |
| 3712 | __m256i __B) { |
| 3713 | return (__m256i)__builtin_ia32_selectd_256(__U, |
| 3714 | (__v8si)_mm256_permutex2var_epi32(__A, __I, __B), |
| 3715 | (__v8si)_mm256_setzero_si256()); |
| 3716 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3717 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3718 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3719 | _mm_permutex2var_pd(__m128d __A, __m128i __I, __m128d __B) { |
| 3720 | return (__m128d)__builtin_ia32_vpermi2varpd128((__v2df)__A, (__v2di)__I, |
| 3721 | (__v2df)__B); |
| 3722 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3723 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3724 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3725 | _mm_mask_permutex2var_pd(__m128d __A, __mmask8 __U, __m128i __I, __m128d __B) { |
| 3726 | return (__m128d)__builtin_ia32_selectpd_128(__U, |
| 3727 | (__v2df)_mm_permutex2var_pd(__A, __I, __B), |
| 3728 | (__v2df)__A); |
| 3729 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3730 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3731 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3732 | _mm_mask2_permutex2var_pd(__m128d __A, __m128i __I, __mmask8 __U, __m128d __B) { |
| 3733 | return (__m128d)__builtin_ia32_selectpd_128(__U, |
| 3734 | (__v2df)_mm_permutex2var_pd(__A, __I, __B), |
| 3735 | (__v2df)(__m128d)__I); |
| 3736 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3737 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3738 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 3739 | _mm_maskz_permutex2var_pd(__mmask8 __U, __m128d __A, __m128i __I, __m128d __B) { |
| 3740 | return (__m128d)__builtin_ia32_selectpd_128(__U, |
| 3741 | (__v2df)_mm_permutex2var_pd(__A, __I, __B), |
| 3742 | (__v2df)_mm_setzero_pd()); |
| 3743 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3744 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3745 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3746 | _mm256_permutex2var_pd(__m256d __A, __m256i __I, __m256d __B) { |
| 3747 | return (__m256d)__builtin_ia32_vpermi2varpd256((__v4df)__A, (__v4di)__I, |
| 3748 | (__v4df)__B); |
| 3749 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3750 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3751 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3752 | _mm256_mask_permutex2var_pd(__m256d __A, __mmask8 __U, __m256i __I, |
| 3753 | __m256d __B) { |
| 3754 | return (__m256d)__builtin_ia32_selectpd_256(__U, |
| 3755 | (__v4df)_mm256_permutex2var_pd(__A, __I, __B), |
| 3756 | (__v4df)__A); |
| 3757 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3758 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3759 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3760 | _mm256_mask2_permutex2var_pd(__m256d __A, __m256i __I, __mmask8 __U, |
| 3761 | __m256d __B) { |
| 3762 | return (__m256d)__builtin_ia32_selectpd_256(__U, |
| 3763 | (__v4df)_mm256_permutex2var_pd(__A, __I, __B), |
| 3764 | (__v4df)(__m256d)__I); |
| 3765 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3766 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3767 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 3768 | _mm256_maskz_permutex2var_pd(__mmask8 __U, __m256d __A, __m256i __I, |
| 3769 | __m256d __B) { |
| 3770 | return (__m256d)__builtin_ia32_selectpd_256(__U, |
| 3771 | (__v4df)_mm256_permutex2var_pd(__A, __I, __B), |
| 3772 | (__v4df)_mm256_setzero_pd()); |
| 3773 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3774 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3775 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3776 | _mm_permutex2var_ps(__m128 __A, __m128i __I, __m128 __B) { |
| 3777 | return (__m128)__builtin_ia32_vpermi2varps128((__v4sf)__A, (__v4si)__I, |
| 3778 | (__v4sf)__B); |
| 3779 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3780 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3781 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3782 | _mm_mask_permutex2var_ps(__m128 __A, __mmask8 __U, __m128i __I, __m128 __B) { |
| 3783 | return (__m128)__builtin_ia32_selectps_128(__U, |
| 3784 | (__v4sf)_mm_permutex2var_ps(__A, __I, __B), |
| 3785 | (__v4sf)__A); |
| 3786 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3787 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3788 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3789 | _mm_mask2_permutex2var_ps(__m128 __A, __m128i __I, __mmask8 __U, __m128 __B) { |
| 3790 | return (__m128)__builtin_ia32_selectps_128(__U, |
| 3791 | (__v4sf)_mm_permutex2var_ps(__A, __I, __B), |
| 3792 | (__v4sf)(__m128)__I); |
| 3793 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3794 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3795 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 3796 | _mm_maskz_permutex2var_ps(__mmask8 __U, __m128 __A, __m128i __I, __m128 __B) { |
| 3797 | return (__m128)__builtin_ia32_selectps_128(__U, |
| 3798 | (__v4sf)_mm_permutex2var_ps(__A, __I, __B), |
| 3799 | (__v4sf)_mm_setzero_ps()); |
| 3800 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3801 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3802 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3803 | _mm256_permutex2var_ps(__m256 __A, __m256i __I, __m256 __B) { |
| 3804 | return (__m256)__builtin_ia32_vpermi2varps256((__v8sf)__A, (__v8si)__I, |
| 3805 | (__v8sf) __B); |
| 3806 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3807 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3808 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3809 | _mm256_mask_permutex2var_ps(__m256 __A, __mmask8 __U, __m256i __I, __m256 __B) { |
| 3810 | return (__m256)__builtin_ia32_selectps_256(__U, |
| 3811 | (__v8sf)_mm256_permutex2var_ps(__A, __I, __B), |
| 3812 | (__v8sf)__A); |
| 3813 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3814 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3815 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3816 | _mm256_mask2_permutex2var_ps(__m256 __A, __m256i __I, __mmask8 __U, |
| 3817 | __m256 __B) { |
| 3818 | return (__m256)__builtin_ia32_selectps_256(__U, |
| 3819 | (__v8sf)_mm256_permutex2var_ps(__A, __I, __B), |
| 3820 | (__v8sf)(__m256)__I); |
| 3821 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3822 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3823 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 3824 | _mm256_maskz_permutex2var_ps(__mmask8 __U, __m256 __A, __m256i __I, |
| 3825 | __m256 __B) { |
| 3826 | return (__m256)__builtin_ia32_selectps_256(__U, |
| 3827 | (__v8sf)_mm256_permutex2var_ps(__A, __I, __B), |
| 3828 | (__v8sf)_mm256_setzero_ps()); |
| 3829 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3830 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3831 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3832 | _mm_permutex2var_epi64(__m128i __A, __m128i __I, __m128i __B) { |
| 3833 | return (__m128i)__builtin_ia32_vpermi2varq128((__v2di)__A, (__v2di)__I, |
| 3834 | (__v2di)__B); |
| 3835 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3836 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3837 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3838 | _mm_mask_permutex2var_epi64(__m128i __A, __mmask8 __U, __m128i __I, |
| 3839 | __m128i __B) { |
| 3840 | return (__m128i)__builtin_ia32_selectq_128(__U, |
| 3841 | (__v2di)_mm_permutex2var_epi64(__A, __I, __B), |
| 3842 | (__v2di)__A); |
| 3843 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3844 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3845 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3846 | _mm_mask2_permutex2var_epi64(__m128i __A, __m128i __I, __mmask8 __U, |
| 3847 | __m128i __B) { |
| 3848 | return (__m128i)__builtin_ia32_selectq_128(__U, |
| 3849 | (__v2di)_mm_permutex2var_epi64(__A, __I, __B), |
| 3850 | (__v2di)__I); |
| 3851 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3852 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3853 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3854 | _mm_maskz_permutex2var_epi64(__mmask8 __U, __m128i __A, __m128i __I, |
| 3855 | __m128i __B) { |
| 3856 | return (__m128i)__builtin_ia32_selectq_128(__U, |
| 3857 | (__v2di)_mm_permutex2var_epi64(__A, __I, __B), |
| 3858 | (__v2di)_mm_setzero_si128()); |
| 3859 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3860 | |
| 3861 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3862 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3863 | _mm256_permutex2var_epi64(__m256i __A, __m256i __I, __m256i __B) { |
| 3864 | return (__m256i)__builtin_ia32_vpermi2varq256((__v4di)__A, (__v4di) __I, |
| 3865 | (__v4di) __B); |
| 3866 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3867 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3868 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3869 | _mm256_mask_permutex2var_epi64(__m256i __A, __mmask8 __U, __m256i __I, |
| 3870 | __m256i __B) { |
| 3871 | return (__m256i)__builtin_ia32_selectq_256(__U, |
| 3872 | (__v4di)_mm256_permutex2var_epi64(__A, __I, __B), |
| 3873 | (__v4di)__A); |
| 3874 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3875 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3876 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3877 | _mm256_mask2_permutex2var_epi64(__m256i __A, __m256i __I, __mmask8 __U, |
| 3878 | __m256i __B) { |
| 3879 | return (__m256i)__builtin_ia32_selectq_256(__U, |
| 3880 | (__v4di)_mm256_permutex2var_epi64(__A, __I, __B), |
| 3881 | (__v4di)__I); |
| 3882 | } |
Craig Topper | 68a272d | 2018-05-29 03:26:38 +0000 | [diff] [blame] | 3883 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3884 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3885 | _mm256_maskz_permutex2var_epi64(__mmask8 __U, __m256i __A, __m256i __I, |
| 3886 | __m256i __B) { |
| 3887 | return (__m256i)__builtin_ia32_selectq_256(__U, |
| 3888 | (__v4di)_mm256_permutex2var_epi64(__A, __I, __B), |
| 3889 | (__v4di)_mm256_setzero_si256()); |
| 3890 | } |
Asaf Badouh | 93aa4c8 | 2015-07-28 12:04:40 +0000 | [diff] [blame] | 3891 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3892 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3893 | _mm_mask_cvtepi8_epi32(__m128i __W, __mmask8 __U, __m128i __A) |
| 3894 | { |
| 3895 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 3896 | (__v4si)_mm_cvtepi8_epi32(__A), |
| 3897 | (__v4si)__W); |
| 3898 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3899 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3900 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3901 | _mm_maskz_cvtepi8_epi32(__mmask8 __U, __m128i __A) |
| 3902 | { |
| 3903 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 3904 | (__v4si)_mm_cvtepi8_epi32(__A), |
| 3905 | (__v4si)_mm_setzero_si128()); |
| 3906 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3907 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3908 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3909 | _mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) |
| 3910 | { |
| 3911 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 3912 | (__v8si)_mm256_cvtepi8_epi32(__A), |
| 3913 | (__v8si)__W); |
| 3914 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3915 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3916 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3917 | _mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) |
| 3918 | { |
| 3919 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 3920 | (__v8si)_mm256_cvtepi8_epi32(__A), |
| 3921 | (__v8si)_mm256_setzero_si256()); |
| 3922 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3923 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3924 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3925 | _mm_mask_cvtepi8_epi64(__m128i __W, __mmask8 __U, __m128i __A) |
| 3926 | { |
| 3927 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 3928 | (__v2di)_mm_cvtepi8_epi64(__A), |
| 3929 | (__v2di)__W); |
| 3930 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3931 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3932 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3933 | _mm_maskz_cvtepi8_epi64(__mmask8 __U, __m128i __A) |
| 3934 | { |
| 3935 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 3936 | (__v2di)_mm_cvtepi8_epi64(__A), |
| 3937 | (__v2di)_mm_setzero_si128()); |
| 3938 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3939 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3940 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3941 | _mm256_mask_cvtepi8_epi64(__m256i __W, __mmask8 __U, __m128i __A) |
| 3942 | { |
| 3943 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 3944 | (__v4di)_mm256_cvtepi8_epi64(__A), |
| 3945 | (__v4di)__W); |
| 3946 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3947 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3948 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3949 | _mm256_maskz_cvtepi8_epi64(__mmask8 __U, __m128i __A) |
| 3950 | { |
| 3951 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 3952 | (__v4di)_mm256_cvtepi8_epi64(__A), |
| 3953 | (__v4di)_mm256_setzero_si256()); |
| 3954 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3955 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3956 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3957 | _mm_mask_cvtepi32_epi64(__m128i __W, __mmask8 __U, __m128i __X) |
| 3958 | { |
| 3959 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 3960 | (__v2di)_mm_cvtepi32_epi64(__X), |
| 3961 | (__v2di)__W); |
| 3962 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3963 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3964 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3965 | _mm_maskz_cvtepi32_epi64(__mmask8 __U, __m128i __X) |
| 3966 | { |
| 3967 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 3968 | (__v2di)_mm_cvtepi32_epi64(__X), |
| 3969 | (__v2di)_mm_setzero_si128()); |
| 3970 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3971 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3972 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3973 | _mm256_mask_cvtepi32_epi64(__m256i __W, __mmask8 __U, __m128i __X) |
| 3974 | { |
| 3975 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 3976 | (__v4di)_mm256_cvtepi32_epi64(__X), |
| 3977 | (__v4di)__W); |
| 3978 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3979 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3980 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 3981 | _mm256_maskz_cvtepi32_epi64(__mmask8 __U, __m128i __X) |
| 3982 | { |
| 3983 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 3984 | (__v4di)_mm256_cvtepi32_epi64(__X), |
| 3985 | (__v4di)_mm256_setzero_si256()); |
| 3986 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3987 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3988 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3989 | _mm_mask_cvtepi16_epi32(__m128i __W, __mmask8 __U, __m128i __A) |
| 3990 | { |
| 3991 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 3992 | (__v4si)_mm_cvtepi16_epi32(__A), |
| 3993 | (__v4si)__W); |
| 3994 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 3995 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 3996 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 3997 | _mm_maskz_cvtepi16_epi32(__mmask8 __U, __m128i __A) |
| 3998 | { |
| 3999 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4000 | (__v4si)_mm_cvtepi16_epi32(__A), |
| 4001 | (__v4si)_mm_setzero_si128()); |
| 4002 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4003 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4004 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4005 | _mm256_mask_cvtepi16_epi32(__m256i __W, __mmask8 __U, __m128i __A) |
| 4006 | { |
| 4007 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4008 | (__v8si)_mm256_cvtepi16_epi32(__A), |
| 4009 | (__v8si)__W); |
| 4010 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4011 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4012 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4013 | _mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) |
| 4014 | { |
| 4015 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4016 | (__v8si)_mm256_cvtepi16_epi32(__A), |
| 4017 | (__v8si)_mm256_setzero_si256()); |
| 4018 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4019 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4020 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4021 | _mm_mask_cvtepi16_epi64(__m128i __W, __mmask8 __U, __m128i __A) |
| 4022 | { |
| 4023 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4024 | (__v2di)_mm_cvtepi16_epi64(__A), |
| 4025 | (__v2di)__W); |
| 4026 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4027 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4028 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4029 | _mm_maskz_cvtepi16_epi64(__mmask8 __U, __m128i __A) |
| 4030 | { |
| 4031 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4032 | (__v2di)_mm_cvtepi16_epi64(__A), |
| 4033 | (__v2di)_mm_setzero_si128()); |
| 4034 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4035 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4036 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4037 | _mm256_mask_cvtepi16_epi64(__m256i __W, __mmask8 __U, __m128i __A) |
| 4038 | { |
| 4039 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4040 | (__v4di)_mm256_cvtepi16_epi64(__A), |
| 4041 | (__v4di)__W); |
| 4042 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4043 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4044 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4045 | _mm256_maskz_cvtepi16_epi64(__mmask8 __U, __m128i __A) |
| 4046 | { |
| 4047 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4048 | (__v4di)_mm256_cvtepi16_epi64(__A), |
| 4049 | (__v4di)_mm256_setzero_si256()); |
| 4050 | } |
Michael Zuckerman | 7cdb72f | 2016-02-18 09:09:34 +0000 | [diff] [blame] | 4051 | |
| 4052 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4053 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4054 | _mm_mask_cvtepu8_epi32(__m128i __W, __mmask8 __U, __m128i __A) |
| 4055 | { |
| 4056 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4057 | (__v4si)_mm_cvtepu8_epi32(__A), |
| 4058 | (__v4si)__W); |
| 4059 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4060 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4061 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4062 | _mm_maskz_cvtepu8_epi32(__mmask8 __U, __m128i __A) |
| 4063 | { |
| 4064 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4065 | (__v4si)_mm_cvtepu8_epi32(__A), |
| 4066 | (__v4si)_mm_setzero_si128()); |
| 4067 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4068 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4069 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4070 | _mm256_mask_cvtepu8_epi32(__m256i __W, __mmask8 __U, __m128i __A) |
| 4071 | { |
| 4072 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4073 | (__v8si)_mm256_cvtepu8_epi32(__A), |
| 4074 | (__v8si)__W); |
| 4075 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4076 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4077 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4078 | _mm256_maskz_cvtepu8_epi32(__mmask8 __U, __m128i __A) |
| 4079 | { |
| 4080 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4081 | (__v8si)_mm256_cvtepu8_epi32(__A), |
| 4082 | (__v8si)_mm256_setzero_si256()); |
| 4083 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4084 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4085 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4086 | _mm_mask_cvtepu8_epi64(__m128i __W, __mmask8 __U, __m128i __A) |
| 4087 | { |
| 4088 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4089 | (__v2di)_mm_cvtepu8_epi64(__A), |
| 4090 | (__v2di)__W); |
| 4091 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4092 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4093 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4094 | _mm_maskz_cvtepu8_epi64(__mmask8 __U, __m128i __A) |
| 4095 | { |
| 4096 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4097 | (__v2di)_mm_cvtepu8_epi64(__A), |
| 4098 | (__v2di)_mm_setzero_si128()); |
| 4099 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4100 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4101 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4102 | _mm256_mask_cvtepu8_epi64(__m256i __W, __mmask8 __U, __m128i __A) |
| 4103 | { |
| 4104 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4105 | (__v4di)_mm256_cvtepu8_epi64(__A), |
| 4106 | (__v4di)__W); |
| 4107 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4108 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4109 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4110 | _mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) |
| 4111 | { |
| 4112 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4113 | (__v4di)_mm256_cvtepu8_epi64(__A), |
| 4114 | (__v4di)_mm256_setzero_si256()); |
| 4115 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4116 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4117 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4118 | _mm_mask_cvtepu32_epi64(__m128i __W, __mmask8 __U, __m128i __X) |
| 4119 | { |
| 4120 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4121 | (__v2di)_mm_cvtepu32_epi64(__X), |
| 4122 | (__v2di)__W); |
| 4123 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4124 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4125 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4126 | _mm_maskz_cvtepu32_epi64(__mmask8 __U, __m128i __X) |
| 4127 | { |
| 4128 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4129 | (__v2di)_mm_cvtepu32_epi64(__X), |
| 4130 | (__v2di)_mm_setzero_si128()); |
| 4131 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4132 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4133 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4134 | _mm256_mask_cvtepu32_epi64(__m256i __W, __mmask8 __U, __m128i __X) |
| 4135 | { |
| 4136 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4137 | (__v4di)_mm256_cvtepu32_epi64(__X), |
| 4138 | (__v4di)__W); |
| 4139 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4140 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4141 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4142 | _mm256_maskz_cvtepu32_epi64(__mmask8 __U, __m128i __X) |
| 4143 | { |
| 4144 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4145 | (__v4di)_mm256_cvtepu32_epi64(__X), |
| 4146 | (__v4di)_mm256_setzero_si256()); |
| 4147 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4148 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4149 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4150 | _mm_mask_cvtepu16_epi32(__m128i __W, __mmask8 __U, __m128i __A) |
| 4151 | { |
| 4152 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4153 | (__v4si)_mm_cvtepu16_epi32(__A), |
| 4154 | (__v4si)__W); |
| 4155 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4156 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4157 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4158 | _mm_maskz_cvtepu16_epi32(__mmask8 __U, __m128i __A) |
| 4159 | { |
| 4160 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4161 | (__v4si)_mm_cvtepu16_epi32(__A), |
| 4162 | (__v4si)_mm_setzero_si128()); |
| 4163 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4164 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4165 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4166 | _mm256_mask_cvtepu16_epi32(__m256i __W, __mmask8 __U, __m128i __A) |
| 4167 | { |
| 4168 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4169 | (__v8si)_mm256_cvtepu16_epi32(__A), |
| 4170 | (__v8si)__W); |
| 4171 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4172 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4173 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4174 | _mm256_maskz_cvtepu16_epi32(__mmask8 __U, __m128i __A) |
| 4175 | { |
| 4176 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4177 | (__v8si)_mm256_cvtepu16_epi32(__A), |
| 4178 | (__v8si)_mm256_setzero_si256()); |
| 4179 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4180 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4181 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4182 | _mm_mask_cvtepu16_epi64(__m128i __W, __mmask8 __U, __m128i __A) |
| 4183 | { |
| 4184 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4185 | (__v2di)_mm_cvtepu16_epi64(__A), |
| 4186 | (__v2di)__W); |
| 4187 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4188 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4189 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4190 | _mm_maskz_cvtepu16_epi64(__mmask8 __U, __m128i __A) |
| 4191 | { |
| 4192 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4193 | (__v2di)_mm_cvtepu16_epi64(__A), |
| 4194 | (__v2di)_mm_setzero_si128()); |
| 4195 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4196 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4197 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4198 | _mm256_mask_cvtepu16_epi64(__m256i __W, __mmask8 __U, __m128i __A) |
| 4199 | { |
| 4200 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4201 | (__v4di)_mm256_cvtepu16_epi64(__A), |
| 4202 | (__v4di)__W); |
| 4203 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4204 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4205 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4206 | _mm256_maskz_cvtepu16_epi64(__mmask8 __U, __m128i __A) |
| 4207 | { |
| 4208 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4209 | (__v4di)_mm256_cvtepu16_epi64(__A), |
| 4210 | (__v4di)_mm256_setzero_si256()); |
| 4211 | } |
Michael Zuckerman | 7a33dce | 2016-02-21 14:00:11 +0000 | [diff] [blame] | 4212 | |
| 4213 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4214 | #define _mm_rol_epi32(a, b) \ |
| 4215 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4216 | (__v4si)_mm_setzero_si128(), \ |
| 4217 | (__mmask8)-1) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4218 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4219 | #define _mm_mask_rol_epi32(w, u, a, b) \ |
| 4220 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4221 | (__v4si)(__m128i)(w), (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4222 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4223 | #define _mm_maskz_rol_epi32(u, a, b) \ |
| 4224 | (__m128i)__builtin_ia32_prold128_mask((__v4si)(__m128i)(a), (int)(b), \ |
| 4225 | (__v4si)_mm_setzero_si128(), \ |
| 4226 | (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4227 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4228 | #define _mm256_rol_epi32(a, b) \ |
| 4229 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4230 | (__v8si)_mm256_setzero_si256(), \ |
| 4231 | (__mmask8)-1) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4232 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4233 | #define _mm256_mask_rol_epi32(w, u, a, b) \ |
| 4234 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4235 | (__v8si)(__m256i)(w), (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4236 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4237 | #define _mm256_maskz_rol_epi32(u, a, b) \ |
| 4238 | (__m256i)__builtin_ia32_prold256_mask((__v8si)(__m256i)(a), (int)(b), \ |
| 4239 | (__v8si)_mm256_setzero_si256(), \ |
| 4240 | (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4241 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4242 | #define _mm_rol_epi64(a, b) \ |
| 4243 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4244 | (__v2di)_mm_setzero_si128(), \ |
| 4245 | (__mmask8)-1) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4246 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4247 | #define _mm_mask_rol_epi64(w, u, a, b) \ |
| 4248 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4249 | (__v2di)(__m128i)(w), (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4250 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4251 | #define _mm_maskz_rol_epi64(u, a, b) \ |
| 4252 | (__m128i)__builtin_ia32_prolq128_mask((__v2di)(__m128i)(a), (int)(b), \ |
| 4253 | (__v2di)_mm_setzero_si128(), \ |
| 4254 | (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4255 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4256 | #define _mm256_rol_epi64(a, b) \ |
| 4257 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4258 | (__v4di)_mm256_setzero_si256(), \ |
| 4259 | (__mmask8)-1) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4260 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4261 | #define _mm256_mask_rol_epi64(w, u, a, b) \ |
| 4262 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4263 | (__v4di)(__m256i)(w), (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4264 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4265 | #define _mm256_maskz_rol_epi64(u, a, b) \ |
| 4266 | (__m256i)__builtin_ia32_prolq256_mask((__v4di)(__m256i)(a), (int)(b), \ |
| 4267 | (__v4di)_mm256_setzero_si256(), \ |
| 4268 | (__mmask8)(u)) |
Michael Zuckerman | 38a2727 | 2016-02-22 09:05:41 +0000 | [diff] [blame] | 4269 | |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4270 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4271 | _mm_rolv_epi32 (__m128i __A, __m128i __B) |
| 4272 | { |
| 4273 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4274 | (__v4si) __B, |
| 4275 | (__v4si) |
| 4276 | _mm_setzero_si128 (), |
| 4277 | (__mmask8) -1); |
| 4278 | } |
| 4279 | |
| 4280 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4281 | _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4282 | __m128i __B) |
| 4283 | { |
| 4284 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4285 | (__v4si) __B, |
| 4286 | (__v4si) __W, |
| 4287 | (__mmask8) __U); |
| 4288 | } |
| 4289 | |
| 4290 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4291 | _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4292 | { |
| 4293 | return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, |
| 4294 | (__v4si) __B, |
| 4295 | (__v4si) |
| 4296 | _mm_setzero_si128 (), |
| 4297 | (__mmask8) __U); |
| 4298 | } |
| 4299 | |
| 4300 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4301 | _mm256_rolv_epi32 (__m256i __A, __m256i __B) |
| 4302 | { |
| 4303 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4304 | (__v8si) __B, |
| 4305 | (__v8si) |
| 4306 | _mm256_setzero_si256 (), |
| 4307 | (__mmask8) -1); |
| 4308 | } |
| 4309 | |
| 4310 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4311 | _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4312 | __m256i __B) |
| 4313 | { |
| 4314 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4315 | (__v8si) __B, |
| 4316 | (__v8si) __W, |
| 4317 | (__mmask8) __U); |
| 4318 | } |
| 4319 | |
| 4320 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4321 | _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4322 | { |
| 4323 | return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, |
| 4324 | (__v8si) __B, |
| 4325 | (__v8si) |
| 4326 | _mm256_setzero_si256 (), |
| 4327 | (__mmask8) __U); |
| 4328 | } |
| 4329 | |
| 4330 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4331 | _mm_rolv_epi64 (__m128i __A, __m128i __B) |
| 4332 | { |
| 4333 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4334 | (__v2di) __B, |
| 4335 | (__v2di) |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4336 | _mm_setzero_si128 (), |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4337 | (__mmask8) -1); |
| 4338 | } |
| 4339 | |
| 4340 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4341 | _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4342 | __m128i __B) |
| 4343 | { |
| 4344 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4345 | (__v2di) __B, |
| 4346 | (__v2di) __W, |
| 4347 | (__mmask8) __U); |
| 4348 | } |
| 4349 | |
| 4350 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4351 | _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4352 | { |
| 4353 | return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, |
| 4354 | (__v2di) __B, |
| 4355 | (__v2di) |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4356 | _mm_setzero_si128 (), |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4357 | (__mmask8) __U); |
| 4358 | } |
| 4359 | |
| 4360 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4361 | _mm256_rolv_epi64 (__m256i __A, __m256i __B) |
| 4362 | { |
| 4363 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4364 | (__v4di) __B, |
| 4365 | (__v4di) |
| 4366 | _mm256_setzero_si256 (), |
| 4367 | (__mmask8) -1); |
| 4368 | } |
| 4369 | |
| 4370 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4371 | _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4372 | __m256i __B) |
| 4373 | { |
| 4374 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4375 | (__v4di) __B, |
| 4376 | (__v4di) __W, |
| 4377 | (__mmask8) __U); |
| 4378 | } |
| 4379 | |
| 4380 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4381 | _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4382 | { |
| 4383 | return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, |
| 4384 | (__v4di) __B, |
| 4385 | (__v4di) |
| 4386 | _mm256_setzero_si256 (), |
| 4387 | (__mmask8) __U); |
| 4388 | } |
| 4389 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4390 | #define _mm_ror_epi32(A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4391 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 4392 | (__v4si)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4393 | (__mmask8)-1) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4394 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4395 | #define _mm_mask_ror_epi32(W, U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4396 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4397 | (__v4si)(__m128i)(W), (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4398 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4399 | #define _mm_maskz_ror_epi32(U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4400 | (__m128i)__builtin_ia32_prord128_mask((__v4si)(__m128i)(A), (int)(B), \ |
| 4401 | (__v4si)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4402 | (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4403 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4404 | #define _mm256_ror_epi32(A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4405 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 4406 | (__v8si)_mm256_setzero_si256(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4407 | (__mmask8)-1) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4408 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4409 | #define _mm256_mask_ror_epi32(W, U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4410 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4411 | (__v8si)(__m256i)(W), (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4412 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4413 | #define _mm256_maskz_ror_epi32(U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4414 | (__m256i)__builtin_ia32_prord256_mask((__v8si)(__m256i)(A), (int)(B), \ |
| 4415 | (__v8si)_mm256_setzero_si256(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4416 | (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4417 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4418 | #define _mm_ror_epi64(A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4419 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4420 | (__v2di)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4421 | (__mmask8)-1) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4422 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4423 | #define _mm_mask_ror_epi64(W, U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4424 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4425 | (__v2di)(__m128i)(W), (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4426 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4427 | #define _mm_maskz_ror_epi64(U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4428 | (__m128i)__builtin_ia32_prorq128_mask((__v2di)(__m128i)(A), (int)(B), \ |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4429 | (__v2di)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4430 | (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4431 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4432 | #define _mm256_ror_epi64(A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4433 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 4434 | (__v4di)_mm256_setzero_si256(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4435 | (__mmask8)-1) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4436 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4437 | #define _mm256_mask_ror_epi64(W, U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4438 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4439 | (__v4di)(__m256i)(W), (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4440 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4441 | #define _mm256_maskz_ror_epi64(U, A, B) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 4442 | (__m256i)__builtin_ia32_prorq256_mask((__v4di)(__m256i)(A), (int)(B), \ |
| 4443 | (__v4di)_mm256_setzero_si256(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 4444 | (__mmask8)(U)) |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4445 | |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4446 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4447 | _mm_mask_sll_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4448 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4449 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4450 | (__v4si)_mm_sll_epi32(__A, __B), |
| 4451 | (__v4si)__W); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4452 | } |
| 4453 | |
| 4454 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4455 | _mm_maskz_sll_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4456 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4457 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4458 | (__v4si)_mm_sll_epi32(__A, __B), |
| 4459 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4460 | } |
| 4461 | |
| 4462 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4463 | _mm256_mask_sll_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4464 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4465 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4466 | (__v8si)_mm256_sll_epi32(__A, __B), |
| 4467 | (__v8si)__W); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4468 | } |
| 4469 | |
| 4470 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4471 | _mm256_maskz_sll_epi32(__mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4472 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4473 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4474 | (__v8si)_mm256_sll_epi32(__A, __B), |
| 4475 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4476 | } |
| 4477 | |
| 4478 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4479 | _mm_mask_slli_epi32(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4480 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4481 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4482 | (__v4si)_mm_slli_epi32(__A, __B), |
| 4483 | (__v4si)__W); |
| 4484 | } |
| 4485 | |
| 4486 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4487 | _mm_maskz_slli_epi32(__mmask8 __U, __m128i __A, int __B) |
| 4488 | { |
| 4489 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4490 | (__v4si)_mm_slli_epi32(__A, __B), |
| 4491 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4492 | } |
| 4493 | |
| 4494 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4495 | _mm256_mask_slli_epi32(__m256i __W, __mmask8 __U, __m256i __A, int __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4496 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4497 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4498 | (__v8si)_mm256_slli_epi32(__A, __B), |
| 4499 | (__v8si)__W); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4500 | } |
| 4501 | |
| 4502 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4503 | _mm256_maskz_slli_epi32(__mmask8 __U, __m256i __A, int __B) |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4504 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4505 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4506 | (__v8si)_mm256_slli_epi32(__A, __B), |
| 4507 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4508 | } |
| 4509 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4510 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4511 | _mm_mask_sll_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 4512 | { |
| 4513 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4514 | (__v2di)_mm_sll_epi64(__A, __B), |
| 4515 | (__v2di)__W); |
| 4516 | } |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4517 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4518 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4519 | _mm_maskz_sll_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 4520 | { |
| 4521 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4522 | (__v2di)_mm_sll_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4523 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4524 | } |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4525 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4526 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4527 | _mm256_mask_sll_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
| 4528 | { |
| 4529 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4530 | (__v4di)_mm256_sll_epi64(__A, __B), |
| 4531 | (__v4di)__W); |
| 4532 | } |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4533 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4534 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4535 | _mm256_maskz_sll_epi64(__mmask8 __U, __m256i __A, __m128i __B) |
| 4536 | { |
| 4537 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4538 | (__v4di)_mm256_sll_epi64(__A, __B), |
| 4539 | (__v4di)_mm256_setzero_si256()); |
| 4540 | } |
Michael Zuckerman | 1ac360c | 2016-03-01 11:38:16 +0000 | [diff] [blame] | 4541 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4542 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4543 | _mm_mask_slli_epi64(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
| 4544 | { |
| 4545 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4546 | (__v2di)_mm_slli_epi64(__A, __B), |
| 4547 | (__v2di)__W); |
| 4548 | } |
| 4549 | |
| 4550 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4551 | _mm_maskz_slli_epi64(__mmask8 __U, __m128i __A, int __B) |
| 4552 | { |
| 4553 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4554 | (__v2di)_mm_slli_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4555 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4556 | } |
| 4557 | |
| 4558 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4559 | _mm256_mask_slli_epi64(__m256i __W, __mmask8 __U, __m256i __A, int __B) |
| 4560 | { |
| 4561 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4562 | (__v4di)_mm256_slli_epi64(__A, __B), |
| 4563 | (__v4di)__W); |
| 4564 | } |
| 4565 | |
| 4566 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4567 | _mm256_maskz_slli_epi64(__mmask8 __U, __m256i __A, int __B) |
| 4568 | { |
| 4569 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4570 | (__v4di)_mm256_slli_epi64(__A, __B), |
| 4571 | (__v4di)_mm256_setzero_si256()); |
| 4572 | } |
Michael Zuckerman | 0231f16 | 2016-02-23 13:41:13 +0000 | [diff] [blame] | 4573 | |
Michael Zuckerman | e98cc74 | 2016-02-23 15:59:47 +0000 | [diff] [blame] | 4574 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4575 | _mm_rorv_epi32 (__m128i __A, __m128i __B) |
| 4576 | { |
| 4577 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 4578 | (__v4si) __B, |
| 4579 | (__v4si) |
| 4580 | _mm_setzero_si128 (), |
| 4581 | (__mmask8) -1); |
| 4582 | } |
| 4583 | |
| 4584 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4585 | _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4586 | __m128i __B) |
| 4587 | { |
| 4588 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 4589 | (__v4si) __B, |
| 4590 | (__v4si) __W, |
| 4591 | (__mmask8) __U); |
| 4592 | } |
| 4593 | |
| 4594 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4595 | _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4596 | { |
| 4597 | return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, |
| 4598 | (__v4si) __B, |
| 4599 | (__v4si) |
| 4600 | _mm_setzero_si128 (), |
| 4601 | (__mmask8) __U); |
| 4602 | } |
| 4603 | |
| 4604 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4605 | _mm256_rorv_epi32 (__m256i __A, __m256i __B) |
| 4606 | { |
| 4607 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 4608 | (__v8si) __B, |
| 4609 | (__v8si) |
| 4610 | _mm256_setzero_si256 (), |
| 4611 | (__mmask8) -1); |
| 4612 | } |
| 4613 | |
| 4614 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4615 | _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4616 | __m256i __B) |
| 4617 | { |
| 4618 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 4619 | (__v8si) __B, |
| 4620 | (__v8si) __W, |
| 4621 | (__mmask8) __U); |
| 4622 | } |
| 4623 | |
| 4624 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4625 | _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4626 | { |
| 4627 | return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, |
| 4628 | (__v8si) __B, |
| 4629 | (__v8si) |
| 4630 | _mm256_setzero_si256 (), |
| 4631 | (__mmask8) __U); |
| 4632 | } |
| 4633 | |
| 4634 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4635 | _mm_rorv_epi64 (__m128i __A, __m128i __B) |
| 4636 | { |
| 4637 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 4638 | (__v2di) __B, |
| 4639 | (__v2di) |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4640 | _mm_setzero_si128 (), |
Michael Zuckerman | e98cc74 | 2016-02-23 15:59:47 +0000 | [diff] [blame] | 4641 | (__mmask8) -1); |
| 4642 | } |
| 4643 | |
| 4644 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4645 | _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, |
| 4646 | __m128i __B) |
| 4647 | { |
| 4648 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 4649 | (__v2di) __B, |
| 4650 | (__v2di) __W, |
| 4651 | (__mmask8) __U); |
| 4652 | } |
| 4653 | |
| 4654 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4655 | _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) |
| 4656 | { |
| 4657 | return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, |
| 4658 | (__v2di) __B, |
| 4659 | (__v2di) |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4660 | _mm_setzero_si128 (), |
Michael Zuckerman | e98cc74 | 2016-02-23 15:59:47 +0000 | [diff] [blame] | 4661 | (__mmask8) __U); |
| 4662 | } |
| 4663 | |
| 4664 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4665 | _mm256_rorv_epi64 (__m256i __A, __m256i __B) |
| 4666 | { |
| 4667 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 4668 | (__v4di) __B, |
| 4669 | (__v4di) |
| 4670 | _mm256_setzero_si256 (), |
| 4671 | (__mmask8) -1); |
| 4672 | } |
| 4673 | |
| 4674 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4675 | _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, |
| 4676 | __m256i __B) |
| 4677 | { |
| 4678 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 4679 | (__v4di) __B, |
| 4680 | (__v4di) __W, |
| 4681 | (__mmask8) __U); |
| 4682 | } |
| 4683 | |
| 4684 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4685 | _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) |
| 4686 | { |
| 4687 | return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, |
| 4688 | (__v4di) __B, |
| 4689 | (__v4di) |
| 4690 | _mm256_setzero_si256 (), |
| 4691 | (__mmask8) __U); |
| 4692 | } |
| 4693 | |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4694 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4695 | _mm_mask_sllv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4696 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4697 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4698 | (__v2di)_mm_sllv_epi64(__X, __Y), |
| 4699 | (__v2di)__W); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4700 | } |
| 4701 | |
| 4702 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4703 | _mm_maskz_sllv_epi64(__mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4704 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4705 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4706 | (__v2di)_mm_sllv_epi64(__X, __Y), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4707 | (__v2di)_mm_setzero_si128()); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4708 | } |
| 4709 | |
| 4710 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4711 | _mm256_mask_sllv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4712 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4713 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4714 | (__v4di)_mm256_sllv_epi64(__X, __Y), |
| 4715 | (__v4di)__W); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4716 | } |
| 4717 | |
| 4718 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4719 | _mm256_maskz_sllv_epi64(__mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4720 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4721 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4722 | (__v4di)_mm256_sllv_epi64(__X, __Y), |
| 4723 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4724 | } |
| 4725 | |
| 4726 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4727 | _mm_mask_sllv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4728 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4729 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4730 | (__v4si)_mm_sllv_epi32(__X, __Y), |
| 4731 | (__v4si)__W); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4732 | } |
| 4733 | |
| 4734 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4735 | _mm_maskz_sllv_epi32(__mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4736 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4737 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4738 | (__v4si)_mm_sllv_epi32(__X, __Y), |
| 4739 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4740 | } |
| 4741 | |
| 4742 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4743 | _mm256_mask_sllv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4744 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4745 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4746 | (__v8si)_mm256_sllv_epi32(__X, __Y), |
| 4747 | (__v8si)__W); |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4748 | } |
| 4749 | |
| 4750 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4751 | _mm256_maskz_sllv_epi32(__mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | 431b0e1 | 2016-02-28 07:39:34 +0000 | [diff] [blame] | 4752 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4753 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4754 | (__v8si)_mm256_sllv_epi32(__X, __Y), |
| 4755 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4756 | } |
| 4757 | |
| 4758 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4759 | _mm_mask_srlv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4760 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4761 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4762 | (__v2di)_mm_srlv_epi64(__X, __Y), |
| 4763 | (__v2di)__W); |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4764 | } |
| 4765 | |
| 4766 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4767 | _mm_maskz_srlv_epi64(__mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4768 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4769 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4770 | (__v2di)_mm_srlv_epi64(__X, __Y), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4771 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4772 | } |
| 4773 | |
| 4774 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4775 | _mm256_mask_srlv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
| 4776 | { |
| 4777 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4778 | (__v4di)_mm256_srlv_epi64(__X, __Y), |
| 4779 | (__v4di)__W); |
| 4780 | } |
| 4781 | |
| 4782 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4783 | _mm256_maskz_srlv_epi64(__mmask8 __U, __m256i __X, __m256i __Y) |
| 4784 | { |
| 4785 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4786 | (__v4di)_mm256_srlv_epi64(__X, __Y), |
| 4787 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4788 | } |
| 4789 | |
| 4790 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4791 | _mm_mask_srlv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | 0165e76 | 2016-03-01 13:03:45 +0000 | [diff] [blame] | 4792 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4793 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4794 | (__v4si)_mm_srlv_epi32(__X, __Y), |
| 4795 | (__v4si)__W); |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4796 | } |
| 4797 | |
| 4798 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4799 | _mm_maskz_srlv_epi32(__mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4800 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4801 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4802 | (__v4si)_mm_srlv_epi32(__X, __Y), |
| 4803 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4804 | } |
| 4805 | |
| 4806 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4807 | _mm256_mask_srlv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4808 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4809 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4810 | (__v8si)_mm256_srlv_epi32(__X, __Y), |
| 4811 | (__v8si)__W); |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4812 | } |
| 4813 | |
| 4814 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4815 | _mm256_maskz_srlv_epi32(__mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4816 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4817 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4818 | (__v8si)_mm256_srlv_epi32(__X, __Y), |
| 4819 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4820 | } |
| 4821 | |
| 4822 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4823 | _mm_mask_srl_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | d176d74 | 2016-03-01 17:49:03 +0000 | [diff] [blame] | 4824 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4825 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4826 | (__v4si)_mm_srl_epi32(__A, __B), |
| 4827 | (__v4si)__W); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4828 | } |
| 4829 | |
| 4830 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4831 | _mm_maskz_srl_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4832 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4833 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4834 | (__v4si)_mm_srl_epi32(__A, __B), |
| 4835 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4836 | } |
| 4837 | |
| 4838 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4839 | _mm256_mask_srl_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4840 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4841 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4842 | (__v8si)_mm256_srl_epi32(__A, __B), |
| 4843 | (__v8si)__W); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4844 | } |
| 4845 | |
| 4846 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4847 | _mm256_maskz_srl_epi32(__mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4848 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4849 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4850 | (__v8si)_mm256_srl_epi32(__A, __B), |
| 4851 | (__v8si)_mm256_setzero_si256()); |
| 4852 | } |
| 4853 | |
| 4854 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4855 | _mm_mask_srli_epi32(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
| 4856 | { |
| 4857 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4858 | (__v4si)_mm_srli_epi32(__A, __B), |
| 4859 | (__v4si)__W); |
| 4860 | } |
| 4861 | |
| 4862 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4863 | _mm_maskz_srli_epi32(__mmask8 __U, __m128i __A, int __B) |
| 4864 | { |
| 4865 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4866 | (__v4si)_mm_srli_epi32(__A, __B), |
| 4867 | (__v4si)_mm_setzero_si128()); |
| 4868 | } |
| 4869 | |
| 4870 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4871 | _mm256_mask_srli_epi32(__m256i __W, __mmask8 __U, __m256i __A, int __B) |
| 4872 | { |
| 4873 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4874 | (__v8si)_mm256_srli_epi32(__A, __B), |
| 4875 | (__v8si)__W); |
| 4876 | } |
| 4877 | |
| 4878 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4879 | _mm256_maskz_srli_epi32(__mmask8 __U, __m256i __A, int __B) |
| 4880 | { |
| 4881 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4882 | (__v8si)_mm256_srli_epi32(__A, __B), |
| 4883 | (__v8si)_mm256_setzero_si256()); |
| 4884 | } |
| 4885 | |
| 4886 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4887 | _mm_mask_srl_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
| 4888 | { |
| 4889 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4890 | (__v2di)_mm_srl_epi64(__A, __B), |
| 4891 | (__v2di)__W); |
| 4892 | } |
| 4893 | |
| 4894 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4895 | _mm_maskz_srl_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
| 4896 | { |
| 4897 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4898 | (__v2di)_mm_srl_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4899 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4900 | } |
| 4901 | |
| 4902 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4903 | _mm256_mask_srl_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
| 4904 | { |
| 4905 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4906 | (__v4di)_mm256_srl_epi64(__A, __B), |
| 4907 | (__v4di)__W); |
| 4908 | } |
| 4909 | |
| 4910 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4911 | _mm256_maskz_srl_epi64(__mmask8 __U, __m256i __A, __m128i __B) |
| 4912 | { |
| 4913 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4914 | (__v4di)_mm256_srl_epi64(__A, __B), |
| 4915 | (__v4di)_mm256_setzero_si256()); |
| 4916 | } |
| 4917 | |
| 4918 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4919 | _mm_mask_srli_epi64(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
| 4920 | { |
| 4921 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4922 | (__v2di)_mm_srli_epi64(__A, __B), |
| 4923 | (__v2di)__W); |
| 4924 | } |
| 4925 | |
| 4926 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4927 | _mm_maskz_srli_epi64(__mmask8 __U, __m128i __A, int __B) |
| 4928 | { |
| 4929 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4930 | (__v2di)_mm_srli_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 4931 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 4932 | } |
| 4933 | |
| 4934 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4935 | _mm256_mask_srli_epi64(__m256i __W, __mmask8 __U, __m256i __A, int __B) |
| 4936 | { |
| 4937 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4938 | (__v4di)_mm256_srli_epi64(__A, __B), |
| 4939 | (__v4di)__W); |
| 4940 | } |
| 4941 | |
| 4942 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4943 | _mm256_maskz_srli_epi64(__mmask8 __U, __m256i __A, int __B) |
| 4944 | { |
| 4945 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 4946 | (__v4di)_mm256_srli_epi64(__A, __B), |
| 4947 | (__v4di)_mm256_setzero_si256()); |
| 4948 | } |
| 4949 | |
| 4950 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4951 | _mm_mask_srav_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
| 4952 | { |
| 4953 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4954 | (__v4si)_mm_srav_epi32(__X, __Y), |
| 4955 | (__v4si)__W); |
| 4956 | } |
| 4957 | |
| 4958 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 4959 | _mm_maskz_srav_epi32(__mmask8 __U, __m128i __X, __m128i __Y) |
| 4960 | { |
| 4961 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 4962 | (__v4si)_mm_srav_epi32(__X, __Y), |
| 4963 | (__v4si)_mm_setzero_si128()); |
| 4964 | } |
| 4965 | |
| 4966 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4967 | _mm256_mask_srav_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
| 4968 | { |
| 4969 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4970 | (__v8si)_mm256_srav_epi32(__X, __Y), |
| 4971 | (__v8si)__W); |
| 4972 | } |
| 4973 | |
| 4974 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 4975 | _mm256_maskz_srav_epi32(__mmask8 __U, __m256i __X, __m256i __Y) |
| 4976 | { |
| 4977 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 4978 | (__v8si)_mm256_srav_epi32(__X, __Y), |
| 4979 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4980 | } |
| 4981 | |
| 4982 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4983 | _mm_srav_epi64(__m128i __X, __m128i __Y) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4984 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4985 | return (__m128i)__builtin_ia32_psravq128((__v2di)__X, (__v2di)__Y); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4986 | } |
| 4987 | |
| 4988 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4989 | _mm_mask_srav_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4990 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4991 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 4992 | (__v2di)_mm_srav_epi64(__X, __Y), |
| 4993 | (__v2di)__W); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4994 | } |
| 4995 | |
| 4996 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4997 | _mm_maskz_srav_epi64(__mmask8 __U, __m128i __X, __m128i __Y) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 4998 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 4999 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 5000 | (__v2di)_mm_srav_epi64(__X, __Y), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 5001 | (__v2di)_mm_setzero_si128()); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5002 | } |
| 5003 | |
| 5004 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 5005 | _mm256_srav_epi64(__m256i __X, __m256i __Y) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5006 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 5007 | return (__m256i)__builtin_ia32_psravq256((__v4di)__X, (__v4di) __Y); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5008 | } |
| 5009 | |
| 5010 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 5011 | _mm256_mask_srav_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5012 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 5013 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 5014 | (__v4di)_mm256_srav_epi64(__X, __Y), |
| 5015 | (__v4di)__W); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5016 | } |
| 5017 | |
| 5018 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5019 | _mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) |
| 5020 | { |
Craig Topper | 5e0709d | 2016-11-13 07:26:34 +0000 | [diff] [blame] | 5021 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 5022 | (__v4di)_mm256_srav_epi64(__X, __Y), |
| 5023 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | d15c95a | 2016-03-02 09:05:46 +0000 | [diff] [blame] | 5024 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5025 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5026 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5027 | _mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5028 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5029 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5030 | (__v4si) __A, |
| 5031 | (__v4si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5032 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5033 | |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5034 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5035 | _mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A) |
| 5036 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5037 | return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U, |
| 5038 | (__v4si) __A, |
| 5039 | (__v4si) _mm_setzero_si128 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5040 | } |
| 5041 | |
| 5042 | |
| 5043 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5044 | _mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5045 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5046 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5047 | (__v8si) __A, |
| 5048 | (__v8si) __W); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5049 | } |
| 5050 | |
| 5051 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5052 | _mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A) |
| 5053 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5054 | return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U, |
| 5055 | (__v8si) __A, |
| 5056 | (__v8si) _mm256_setzero_si256 ()); |
Michael Zuckerman | e654200 | 2016-05-23 08:01:48 +0000 | [diff] [blame] | 5057 | } |
| 5058 | |
| 5059 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5060 | _mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 5061 | { |
| 5062 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5063 | (__v4si) __W, |
| 5064 | (__mmask8) |
| 5065 | __U); |
| 5066 | } |
| 5067 | |
| 5068 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5069 | _mm_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5070 | { |
| 5071 | return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P, |
| 5072 | (__v4si) |
| 5073 | _mm_setzero_si128 (), |
| 5074 | (__mmask8) |
| 5075 | __U); |
| 5076 | } |
| 5077 | |
| 5078 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5079 | _mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 5080 | { |
| 5081 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5082 | (__v8si) __W, |
| 5083 | (__mmask8) |
| 5084 | __U); |
| 5085 | } |
| 5086 | |
| 5087 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5088 | _mm256_maskz_load_epi32 (__mmask8 __U, void const *__P) |
| 5089 | { |
| 5090 | return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P, |
| 5091 | (__v8si) |
| 5092 | _mm256_setzero_si256 (), |
| 5093 | (__mmask8) |
| 5094 | __U); |
| 5095 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5096 | |
| 5097 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5098 | _mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 5099 | { |
| 5100 | __builtin_ia32_movdqa32store128_mask ((__v4si *) __P, |
| 5101 | (__v4si) __A, |
| 5102 | (__mmask8) __U); |
| 5103 | } |
| 5104 | |
| 5105 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5106 | _mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 5107 | { |
| 5108 | __builtin_ia32_movdqa32store256_mask ((__v8si *) __P, |
| 5109 | (__v8si) __A, |
| 5110 | (__mmask8) __U); |
| 5111 | } |
| 5112 | |
| 5113 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5114 | _mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A) |
| 5115 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5116 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5117 | (__v2di) __A, |
| 5118 | (__v2di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5119 | } |
| 5120 | |
| 5121 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5122 | _mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A) |
| 5123 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5124 | return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U, |
| 5125 | (__v2di) __A, |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 5126 | (__v2di) _mm_setzero_si128 ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5127 | } |
| 5128 | |
| 5129 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5130 | _mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A) |
| 5131 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5132 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5133 | (__v4di) __A, |
| 5134 | (__v4di) __W); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5135 | } |
| 5136 | |
| 5137 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5138 | _mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A) |
| 5139 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 5140 | return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U, |
| 5141 | (__v4di) __A, |
| 5142 | (__v4di) _mm256_setzero_si256 ()); |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5143 | } |
| 5144 | |
| 5145 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5146 | _mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 5147 | { |
| 5148 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5149 | (__v2di) __W, |
| 5150 | (__mmask8) |
| 5151 | __U); |
| 5152 | } |
| 5153 | |
| 5154 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5155 | _mm_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5156 | { |
| 5157 | return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P, |
| 5158 | (__v2di) |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 5159 | _mm_setzero_si128 (), |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5160 | (__mmask8) |
| 5161 | __U); |
| 5162 | } |
| 5163 | |
| 5164 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5165 | _mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 5166 | { |
| 5167 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5168 | (__v4di) __W, |
| 5169 | (__mmask8) |
| 5170 | __U); |
| 5171 | } |
| 5172 | |
| 5173 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5174 | _mm256_maskz_load_epi64 (__mmask8 __U, void const *__P) |
| 5175 | { |
| 5176 | return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P, |
| 5177 | (__v4di) |
| 5178 | _mm256_setzero_si256 (), |
| 5179 | (__mmask8) |
| 5180 | __U); |
| 5181 | } |
| 5182 | |
| 5183 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5184 | _mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 5185 | { |
| 5186 | __builtin_ia32_movdqa64store128_mask ((__v2di *) __P, |
| 5187 | (__v2di) __A, |
| 5188 | (__mmask8) __U); |
| 5189 | } |
| 5190 | |
| 5191 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5192 | _mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 5193 | { |
| 5194 | __builtin_ia32_movdqa64store256_mask ((__v4di *) __P, |
| 5195 | (__v4di) __A, |
| 5196 | (__mmask8) __U); |
| 5197 | } |
| 5198 | |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5199 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5200 | _mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 5201 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5202 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5203 | (__v2df)_mm_movedup_pd(__A), |
| 5204 | (__v2df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5205 | } |
| 5206 | |
| 5207 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5208 | _mm_maskz_movedup_pd (__mmask8 __U, __m128d __A) |
| 5209 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5210 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5211 | (__v2df)_mm_movedup_pd(__A), |
| 5212 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5213 | } |
| 5214 | |
| 5215 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5216 | _mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 5217 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5218 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5219 | (__v4df)_mm256_movedup_pd(__A), |
| 5220 | (__v4df)__W); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5221 | } |
| 5222 | |
| 5223 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5224 | _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A) |
| 5225 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 5226 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5227 | (__v4df)_mm256_movedup_pd(__A), |
| 5228 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | 0d67e4b | 2016-03-03 13:43:05 +0000 | [diff] [blame] | 5229 | } |
Michael Zuckerman | ffbb67a | 2016-03-03 09:26:01 +0000 | [diff] [blame] | 5230 | |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5231 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5232 | _mm_mask_set1_epi32(__m128i __O, __mmask8 __M, int __A) |
| 5233 | { |
| 5234 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 5235 | (__v4si) _mm_set1_epi32(__A), |
| 5236 | (__v4si)__O); |
| 5237 | } |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5238 | |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5239 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5240 | _mm_maskz_set1_epi32( __mmask8 __M, int __A) |
| 5241 | { |
| 5242 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 5243 | (__v4si) _mm_set1_epi32(__A), |
| 5244 | (__v4si)_mm_setzero_si128()); |
| 5245 | } |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5246 | |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5247 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5248 | _mm256_mask_set1_epi32(__m256i __O, __mmask8 __M, int __A) |
| 5249 | { |
| 5250 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 5251 | (__v8si) _mm256_set1_epi32(__A), |
| 5252 | (__v8si)__O); |
| 5253 | } |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5254 | |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5255 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5256 | _mm256_maskz_set1_epi32( __mmask8 __M, int __A) |
| 5257 | { |
| 5258 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 5259 | (__v8si) _mm256_set1_epi32(__A), |
| 5260 | (__v8si)_mm256_setzero_si256()); |
| 5261 | } |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5262 | |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5263 | |
| 5264 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5265 | _mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A) |
| 5266 | { |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5267 | return (__m128i) __builtin_ia32_selectq_128(__M, |
Jina Nahias | 123c599 | 2017-09-25 13:38:08 +0000 | [diff] [blame] | 5268 | (__v2di) _mm_set1_epi64x(__A), |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5269 | (__v2di) __O); |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5270 | } |
| 5271 | |
| 5272 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5273 | _mm_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 5274 | { |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5275 | return (__m128i) __builtin_ia32_selectq_128(__M, |
Jina Nahias | 123c599 | 2017-09-25 13:38:08 +0000 | [diff] [blame] | 5276 | (__v2di) _mm_set1_epi64x(__A), |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5277 | (__v2di) _mm_setzero_si128()); |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5278 | } |
| 5279 | |
| 5280 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5281 | _mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A) |
| 5282 | { |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5283 | return (__m256i) __builtin_ia32_selectq_256(__M, |
| 5284 | (__v4di) _mm256_set1_epi64x(__A), |
| 5285 | (__v4di) __O) ; |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5286 | } |
| 5287 | |
| 5288 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5289 | _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A) |
| 5290 | { |
Jina Nahias | 3ad702a | 2017-09-19 11:00:27 +0000 | [diff] [blame] | 5291 | return (__m256i) __builtin_ia32_selectq_256(__M, |
| 5292 | (__v4di) _mm256_set1_epi64x(__A), |
| 5293 | (__v4di) _mm256_setzero_si256()); |
Michael Zuckerman | 912be16 | 2016-03-07 08:29:10 +0000 | [diff] [blame] | 5294 | } |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5295 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5296 | #define _mm_fixupimm_pd(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5297 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 5298 | (__v2df)(__m128d)(B), \ |
| 5299 | (__v2di)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5300 | (__mmask8)-1) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5301 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5302 | #define _mm_mask_fixupimm_pd(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5303 | (__m128d)__builtin_ia32_fixupimmpd128_mask((__v2df)(__m128d)(A), \ |
| 5304 | (__v2df)(__m128d)(B), \ |
| 5305 | (__v2di)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5306 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5307 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5308 | #define _mm_maskz_fixupimm_pd(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5309 | (__m128d)__builtin_ia32_fixupimmpd128_maskz((__v2df)(__m128d)(A), \ |
| 5310 | (__v2df)(__m128d)(B), \ |
| 5311 | (__v2di)(__m128i)(C), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5312 | (int)(imm), (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5313 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5314 | #define _mm256_fixupimm_pd(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5315 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 5316 | (__v4df)(__m256d)(B), \ |
| 5317 | (__v4di)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5318 | (__mmask8)-1) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5319 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5320 | #define _mm256_mask_fixupimm_pd(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5321 | (__m256d)__builtin_ia32_fixupimmpd256_mask((__v4df)(__m256d)(A), \ |
| 5322 | (__v4df)(__m256d)(B), \ |
| 5323 | (__v4di)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5324 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5325 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5326 | #define _mm256_maskz_fixupimm_pd(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5327 | (__m256d)__builtin_ia32_fixupimmpd256_maskz((__v4df)(__m256d)(A), \ |
| 5328 | (__v4df)(__m256d)(B), \ |
| 5329 | (__v4di)(__m256i)(C), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5330 | (int)(imm), (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5331 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5332 | #define _mm_fixupimm_ps(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5333 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 5334 | (__v4sf)(__m128)(B), \ |
| 5335 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5336 | (__mmask8)-1) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5337 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5338 | #define _mm_mask_fixupimm_ps(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5339 | (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \ |
| 5340 | (__v4sf)(__m128)(B), \ |
| 5341 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5342 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5343 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5344 | #define _mm_maskz_fixupimm_ps(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5345 | (__m128)__builtin_ia32_fixupimmps128_maskz((__v4sf)(__m128)(A), \ |
| 5346 | (__v4sf)(__m128)(B), \ |
| 5347 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5348 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5349 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5350 | #define _mm256_fixupimm_ps(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5351 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 5352 | (__v8sf)(__m256)(B), \ |
| 5353 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5354 | (__mmask8)-1) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5355 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5356 | #define _mm256_mask_fixupimm_ps(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5357 | (__m256)__builtin_ia32_fixupimmps256_mask((__v8sf)(__m256)(A), \ |
| 5358 | (__v8sf)(__m256)(B), \ |
| 5359 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5360 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5361 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5362 | #define _mm256_maskz_fixupimm_ps(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 5363 | (__m256)__builtin_ia32_fixupimmps256_maskz((__v8sf)(__m256)(A), \ |
| 5364 | (__v8sf)(__m256)(B), \ |
| 5365 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5366 | (__mmask8)(U)) |
Michael Zuckerman | def7875 | 2016-03-28 12:23:09 +0000 | [diff] [blame] | 5367 | |
| 5368 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5369 | _mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 5370 | { |
| 5371 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 5372 | (__v2df) __W, |
| 5373 | (__mmask8) __U); |
| 5374 | } |
| 5375 | |
| 5376 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5377 | _mm_maskz_load_pd (__mmask8 __U, void const *__P) |
| 5378 | { |
| 5379 | return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P, |
| 5380 | (__v2df) |
| 5381 | _mm_setzero_pd (), |
| 5382 | (__mmask8) __U); |
| 5383 | } |
| 5384 | |
| 5385 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5386 | _mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 5387 | { |
| 5388 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 5389 | (__v4df) __W, |
| 5390 | (__mmask8) __U); |
| 5391 | } |
| 5392 | |
| 5393 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5394 | _mm256_maskz_load_pd (__mmask8 __U, void const *__P) |
| 5395 | { |
| 5396 | return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P, |
| 5397 | (__v4df) |
| 5398 | _mm256_setzero_pd (), |
| 5399 | (__mmask8) __U); |
| 5400 | } |
| 5401 | |
| 5402 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5403 | _mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 5404 | { |
| 5405 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 5406 | (__v4sf) __W, |
| 5407 | (__mmask8) __U); |
| 5408 | } |
| 5409 | |
| 5410 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5411 | _mm_maskz_load_ps (__mmask8 __U, void const *__P) |
| 5412 | { |
| 5413 | return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, |
| 5414 | (__v4sf) |
| 5415 | _mm_setzero_ps (), |
| 5416 | (__mmask8) __U); |
| 5417 | } |
| 5418 | |
| 5419 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5420 | _mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 5421 | { |
| 5422 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 5423 | (__v8sf) __W, |
| 5424 | (__mmask8) __U); |
| 5425 | } |
| 5426 | |
| 5427 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5428 | _mm256_maskz_load_ps (__mmask8 __U, void const *__P) |
| 5429 | { |
| 5430 | return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P, |
| 5431 | (__v8sf) |
| 5432 | _mm256_setzero_ps (), |
| 5433 | (__mmask8) __U); |
| 5434 | } |
| 5435 | |
| 5436 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5437 | _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) |
| 5438 | { |
| 5439 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 5440 | (__v2di) __W, |
| 5441 | (__mmask8) __U); |
| 5442 | } |
| 5443 | |
| 5444 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5445 | _mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 5446 | { |
| 5447 | return (__m128i) __builtin_ia32_loaddqudi128_mask ((__v2di *) __P, |
| 5448 | (__v2di) |
| 5449 | _mm_setzero_si128 (), |
| 5450 | (__mmask8) __U); |
| 5451 | } |
| 5452 | |
| 5453 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5454 | _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P) |
| 5455 | { |
| 5456 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 5457 | (__v4di) __W, |
| 5458 | (__mmask8) __U); |
| 5459 | } |
| 5460 | |
| 5461 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5462 | _mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P) |
| 5463 | { |
| 5464 | return (__m256i) __builtin_ia32_loaddqudi256_mask ((__v4di *) __P, |
| 5465 | (__v4di) |
| 5466 | _mm256_setzero_si256 (), |
| 5467 | (__mmask8) __U); |
| 5468 | } |
| 5469 | |
| 5470 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5471 | _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) |
| 5472 | { |
| 5473 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 5474 | (__v4si) __W, |
| 5475 | (__mmask8) __U); |
| 5476 | } |
| 5477 | |
| 5478 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 5479 | _mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 5480 | { |
| 5481 | return (__m128i) __builtin_ia32_loaddqusi128_mask ((__v4si *) __P, |
| 5482 | (__v4si) |
| 5483 | _mm_setzero_si128 (), |
| 5484 | (__mmask8) __U); |
| 5485 | } |
| 5486 | |
| 5487 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5488 | _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P) |
| 5489 | { |
| 5490 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 5491 | (__v8si) __W, |
| 5492 | (__mmask8) __U); |
| 5493 | } |
| 5494 | |
| 5495 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 5496 | _mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P) |
| 5497 | { |
| 5498 | return (__m256i) __builtin_ia32_loaddqusi256_mask ((__v8si *) __P, |
| 5499 | (__v8si) |
| 5500 | _mm256_setzero_si256 (), |
| 5501 | (__mmask8) __U); |
| 5502 | } |
| 5503 | |
| 5504 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5505 | _mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P) |
| 5506 | { |
| 5507 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 5508 | (__v2df) __W, |
| 5509 | (__mmask8) __U); |
| 5510 | } |
| 5511 | |
| 5512 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5513 | _mm_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 5514 | { |
| 5515 | return (__m128d) __builtin_ia32_loadupd128_mask ((__v2df *) __P, |
| 5516 | (__v2df) |
| 5517 | _mm_setzero_pd (), |
| 5518 | (__mmask8) __U); |
| 5519 | } |
| 5520 | |
| 5521 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5522 | _mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P) |
| 5523 | { |
| 5524 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 5525 | (__v4df) __W, |
| 5526 | (__mmask8) __U); |
| 5527 | } |
| 5528 | |
| 5529 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5530 | _mm256_maskz_loadu_pd (__mmask8 __U, void const *__P) |
| 5531 | { |
| 5532 | return (__m256d) __builtin_ia32_loadupd256_mask ((__v4df *) __P, |
| 5533 | (__v4df) |
| 5534 | _mm256_setzero_pd (), |
| 5535 | (__mmask8) __U); |
| 5536 | } |
| 5537 | |
| 5538 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5539 | _mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P) |
| 5540 | { |
| 5541 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 5542 | (__v4sf) __W, |
| 5543 | (__mmask8) __U); |
| 5544 | } |
| 5545 | |
| 5546 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5547 | _mm_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 5548 | { |
| 5549 | return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, |
| 5550 | (__v4sf) |
| 5551 | _mm_setzero_ps (), |
| 5552 | (__mmask8) __U); |
| 5553 | } |
| 5554 | |
| 5555 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5556 | _mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P) |
| 5557 | { |
| 5558 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 5559 | (__v8sf) __W, |
| 5560 | (__mmask8) __U); |
| 5561 | } |
| 5562 | |
| 5563 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5564 | _mm256_maskz_loadu_ps (__mmask8 __U, void const *__P) |
| 5565 | { |
| 5566 | return (__m256) __builtin_ia32_loadups256_mask ((__v8sf *) __P, |
| 5567 | (__v8sf) |
| 5568 | _mm256_setzero_ps (), |
| 5569 | (__mmask8) __U); |
| 5570 | } |
Michael Zuckerman | fa7ccc5 | 2016-04-10 10:51:04 +0000 | [diff] [blame] | 5571 | |
| 5572 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5573 | _mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A) |
| 5574 | { |
| 5575 | __builtin_ia32_storeapd128_mask ((__v2df *) __P, |
| 5576 | (__v2df) __A, |
| 5577 | (__mmask8) __U); |
| 5578 | } |
| 5579 | |
| 5580 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5581 | _mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A) |
| 5582 | { |
| 5583 | __builtin_ia32_storeapd256_mask ((__v4df *) __P, |
| 5584 | (__v4df) __A, |
| 5585 | (__mmask8) __U); |
| 5586 | } |
| 5587 | |
| 5588 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5589 | _mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A) |
| 5590 | { |
| 5591 | __builtin_ia32_storeaps128_mask ((__v4sf *) __P, |
| 5592 | (__v4sf) __A, |
| 5593 | (__mmask8) __U); |
| 5594 | } |
| 5595 | |
| 5596 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5597 | _mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A) |
| 5598 | { |
| 5599 | __builtin_ia32_storeaps256_mask ((__v8sf *) __P, |
| 5600 | (__v8sf) __A, |
| 5601 | (__mmask8) __U); |
| 5602 | } |
| 5603 | |
| 5604 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5605 | _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A) |
| 5606 | { |
| 5607 | __builtin_ia32_storedqudi128_mask ((__v2di *) __P, |
| 5608 | (__v2di) __A, |
| 5609 | (__mmask8) __U); |
| 5610 | } |
| 5611 | |
| 5612 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5613 | _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A) |
| 5614 | { |
| 5615 | __builtin_ia32_storedqudi256_mask ((__v4di *) __P, |
| 5616 | (__v4di) __A, |
| 5617 | (__mmask8) __U); |
| 5618 | } |
| 5619 | |
| 5620 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5621 | _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A) |
| 5622 | { |
| 5623 | __builtin_ia32_storedqusi128_mask ((__v4si *) __P, |
| 5624 | (__v4si) __A, |
| 5625 | (__mmask8) __U); |
| 5626 | } |
| 5627 | |
| 5628 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5629 | _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A) |
| 5630 | { |
| 5631 | __builtin_ia32_storedqusi256_mask ((__v8si *) __P, |
| 5632 | (__v8si) __A, |
| 5633 | (__mmask8) __U); |
| 5634 | } |
| 5635 | |
| 5636 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5637 | _mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A) |
| 5638 | { |
| 5639 | __builtin_ia32_storeupd128_mask ((__v2df *) __P, |
| 5640 | (__v2df) __A, |
| 5641 | (__mmask8) __U); |
| 5642 | } |
| 5643 | |
| 5644 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5645 | _mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A) |
| 5646 | { |
| 5647 | __builtin_ia32_storeupd256_mask ((__v4df *) __P, |
| 5648 | (__v4df) __A, |
| 5649 | (__mmask8) __U); |
| 5650 | } |
| 5651 | |
| 5652 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5653 | _mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A) |
| 5654 | { |
| 5655 | __builtin_ia32_storeups128_mask ((__v4sf *) __P, |
| 5656 | (__v4sf) __A, |
| 5657 | (__mmask8) __U); |
| 5658 | } |
| 5659 | |
| 5660 | static __inline__ void __DEFAULT_FN_ATTRS |
| 5661 | _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A) |
| 5662 | { |
| 5663 | __builtin_ia32_storeups256_mask ((__v8sf *) __P, |
| 5664 | (__v8sf) __A, |
| 5665 | (__mmask8) __U); |
| 5666 | } |
| 5667 | |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5668 | |
| 5669 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5670 | _mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5671 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5672 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5673 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 5674 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5675 | } |
| 5676 | |
| 5677 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5678 | _mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5679 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5680 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5681 | (__v2df)_mm_unpackhi_pd(__A, __B), |
| 5682 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5683 | } |
| 5684 | |
| 5685 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5686 | _mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5687 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5688 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5689 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 5690 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5691 | } |
| 5692 | |
| 5693 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5694 | _mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5695 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5696 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5697 | (__v4df)_mm256_unpackhi_pd(__A, __B), |
| 5698 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5699 | } |
| 5700 | |
| 5701 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5702 | _mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5703 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5704 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5705 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 5706 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5707 | } |
| 5708 | |
| 5709 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5710 | _mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5711 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5712 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5713 | (__v4sf)_mm_unpackhi_ps(__A, __B), |
| 5714 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5715 | } |
| 5716 | |
| 5717 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5718 | _mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5719 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5720 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 5721 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 5722 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5723 | } |
| 5724 | |
| 5725 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5726 | _mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5727 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5728 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 5729 | (__v8sf)_mm256_unpackhi_ps(__A, __B), |
| 5730 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5731 | } |
| 5732 | |
| 5733 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5734 | _mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5735 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5736 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5737 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 5738 | (__v2df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5739 | } |
| 5740 | |
| 5741 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5742 | _mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5743 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5744 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5745 | (__v2df)_mm_unpacklo_pd(__A, __B), |
| 5746 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5747 | } |
| 5748 | |
| 5749 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5750 | _mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5751 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5752 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5753 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 5754 | (__v4df)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5755 | } |
| 5756 | |
| 5757 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5758 | _mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5759 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5760 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5761 | (__v4df)_mm256_unpacklo_pd(__A, __B), |
| 5762 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5763 | } |
| 5764 | |
| 5765 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5766 | _mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5767 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5768 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5769 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 5770 | (__v4sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5771 | } |
| 5772 | |
| 5773 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5774 | _mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5775 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5776 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5777 | (__v4sf)_mm_unpacklo_ps(__A, __B), |
| 5778 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5779 | } |
| 5780 | |
| 5781 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5782 | _mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5783 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5784 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 5785 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 5786 | (__v8sf)__W); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5787 | } |
| 5788 | |
| 5789 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5790 | _mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B) |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5791 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 5792 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 5793 | (__v8sf)_mm256_unpacklo_ps(__A, __B), |
| 5794 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | cdd54c8 | 2016-04-10 12:54:23 +0000 | [diff] [blame] | 5795 | } |
| 5796 | |
| 5797 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5798 | _mm_rcp14_pd (__m128d __A) |
| 5799 | { |
| 5800 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 5801 | (__v2df) |
| 5802 | _mm_setzero_pd (), |
| 5803 | (__mmask8) -1); |
| 5804 | } |
| 5805 | |
| 5806 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5807 | _mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 5808 | { |
| 5809 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 5810 | (__v2df) __W, |
| 5811 | (__mmask8) __U); |
| 5812 | } |
| 5813 | |
| 5814 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 5815 | _mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A) |
| 5816 | { |
| 5817 | return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, |
| 5818 | (__v2df) |
| 5819 | _mm_setzero_pd (), |
| 5820 | (__mmask8) __U); |
| 5821 | } |
| 5822 | |
| 5823 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5824 | _mm256_rcp14_pd (__m256d __A) |
| 5825 | { |
| 5826 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 5827 | (__v4df) |
| 5828 | _mm256_setzero_pd (), |
| 5829 | (__mmask8) -1); |
| 5830 | } |
| 5831 | |
| 5832 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5833 | _mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 5834 | { |
| 5835 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 5836 | (__v4df) __W, |
| 5837 | (__mmask8) __U); |
| 5838 | } |
| 5839 | |
| 5840 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 5841 | _mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A) |
| 5842 | { |
| 5843 | return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, |
| 5844 | (__v4df) |
| 5845 | _mm256_setzero_pd (), |
| 5846 | (__mmask8) __U); |
| 5847 | } |
| 5848 | |
| 5849 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5850 | _mm_rcp14_ps (__m128 __A) |
| 5851 | { |
| 5852 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 5853 | (__v4sf) |
| 5854 | _mm_setzero_ps (), |
| 5855 | (__mmask8) -1); |
| 5856 | } |
| 5857 | |
| 5858 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5859 | _mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 5860 | { |
| 5861 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 5862 | (__v4sf) __W, |
| 5863 | (__mmask8) __U); |
| 5864 | } |
| 5865 | |
| 5866 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 5867 | _mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A) |
| 5868 | { |
| 5869 | return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, |
| 5870 | (__v4sf) |
| 5871 | _mm_setzero_ps (), |
| 5872 | (__mmask8) __U); |
| 5873 | } |
| 5874 | |
| 5875 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5876 | _mm256_rcp14_ps (__m256 __A) |
| 5877 | { |
| 5878 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 5879 | (__v8sf) |
| 5880 | _mm256_setzero_ps (), |
| 5881 | (__mmask8) -1); |
| 5882 | } |
| 5883 | |
| 5884 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5885 | _mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 5886 | { |
| 5887 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 5888 | (__v8sf) __W, |
| 5889 | (__mmask8) __U); |
| 5890 | } |
| 5891 | |
| 5892 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 5893 | _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A) |
| 5894 | { |
| 5895 | return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, |
| 5896 | (__v8sf) |
| 5897 | _mm256_setzero_ps (), |
| 5898 | (__mmask8) __U); |
| 5899 | } |
| 5900 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5901 | #define _mm_mask_permute_pd(W, U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5902 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 5903 | (__v2df)_mm_permute_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5904 | (__v2df)(__m128d)(W)) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5905 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5906 | #define _mm_maskz_permute_pd(U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5907 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 5908 | (__v2df)_mm_permute_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5909 | (__v2df)_mm_setzero_pd()) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5910 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5911 | #define _mm256_mask_permute_pd(W, U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5912 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 5913 | (__v4df)_mm256_permute_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5914 | (__v4df)(__m256d)(W)) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5915 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5916 | #define _mm256_maskz_permute_pd(U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5917 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 5918 | (__v4df)_mm256_permute_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5919 | (__v4df)_mm256_setzero_pd()) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5920 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5921 | #define _mm_mask_permute_ps(W, U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5922 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 5923 | (__v4sf)_mm_permute_ps((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5924 | (__v4sf)(__m128)(W)) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5925 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5926 | #define _mm_maskz_permute_ps(U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5927 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 5928 | (__v4sf)_mm_permute_ps((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5929 | (__v4sf)_mm_setzero_ps()) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5930 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5931 | #define _mm256_mask_permute_ps(W, U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5932 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 5933 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5934 | (__v8sf)(__m256)(W)) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5935 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5936 | #define _mm256_maskz_permute_ps(U, X, C) \ |
Craig Topper | b3a4477 | 2016-07-02 05:36:43 +0000 | [diff] [blame] | 5937 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 5938 | (__v8sf)_mm256_permute_ps((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 5939 | (__v8sf)_mm256_setzero_ps()) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5940 | |
| 5941 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5942 | _mm_mask_permutevar_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5943 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5944 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5945 | (__v2df)_mm_permutevar_pd(__A, __C), |
| 5946 | (__v2df)__W); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5947 | } |
| 5948 | |
| 5949 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5950 | _mm_maskz_permutevar_pd(__mmask8 __U, __m128d __A, __m128i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5951 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5952 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 5953 | (__v2df)_mm_permutevar_pd(__A, __C), |
| 5954 | (__v2df)_mm_setzero_pd()); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5955 | } |
| 5956 | |
| 5957 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5958 | _mm256_mask_permutevar_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5959 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5960 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5961 | (__v4df)_mm256_permutevar_pd(__A, __C), |
| 5962 | (__v4df)__W); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5963 | } |
| 5964 | |
| 5965 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5966 | _mm256_maskz_permutevar_pd(__mmask8 __U, __m256d __A, __m256i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5967 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5968 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 5969 | (__v4df)_mm256_permutevar_pd(__A, __C), |
| 5970 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5971 | } |
| 5972 | |
| 5973 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5974 | _mm_mask_permutevar_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5975 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5976 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5977 | (__v4sf)_mm_permutevar_ps(__A, __C), |
| 5978 | (__v4sf)__W); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5979 | } |
| 5980 | |
| 5981 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5982 | _mm_maskz_permutevar_ps(__mmask8 __U, __m128 __A, __m128i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5983 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5984 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 5985 | (__v4sf)_mm_permutevar_ps(__A, __C), |
| 5986 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5987 | } |
| 5988 | |
| 5989 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5990 | _mm256_mask_permutevar_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5991 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5992 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 5993 | (__v8sf)_mm256_permutevar_ps(__A, __C), |
| 5994 | (__v8sf)__W); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5995 | } |
| 5996 | |
| 5997 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 5998 | _mm256_maskz_permutevar_ps(__mmask8 __U, __m256 __A, __m256i __C) |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 5999 | { |
Craig Topper | 5391c98 | 2016-12-10 20:27:39 +0000 | [diff] [blame] | 6000 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 6001 | (__v8sf)_mm256_permutevar_ps(__A, __C), |
| 6002 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | d8d2f62 | 2016-04-11 07:15:34 +0000 | [diff] [blame] | 6003 | } |
| 6004 | |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6005 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6006 | _mm_test_epi32_mask (__m128i __A, __m128i __B) |
| 6007 | { |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6008 | return _mm_cmpneq_epi32_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6009 | } |
| 6010 | |
| 6011 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6012 | _mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6013 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6014 | return _mm_mask_cmpneq_epi32_mask (__U, _mm_and_si128 (__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6015 | _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6016 | } |
| 6017 | |
| 6018 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6019 | _mm256_test_epi32_mask (__m256i __A, __m256i __B) |
| 6020 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6021 | return _mm256_cmpneq_epi32_mask (_mm256_and_si256 (__A, __B), |
| 6022 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6023 | } |
| 6024 | |
| 6025 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6026 | _mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6027 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6028 | return _mm256_mask_cmpneq_epi32_mask (__U, _mm256_and_si256 (__A, __B), |
| 6029 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6030 | } |
| 6031 | |
| 6032 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6033 | _mm_test_epi64_mask (__m128i __A, __m128i __B) |
| 6034 | { |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6035 | return _mm_cmpneq_epi64_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6036 | } |
| 6037 | |
| 6038 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6039 | _mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6040 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6041 | return _mm_mask_cmpneq_epi64_mask (__U, _mm_and_si128 (__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6042 | _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6043 | } |
| 6044 | |
| 6045 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6046 | _mm256_test_epi64_mask (__m256i __A, __m256i __B) |
| 6047 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6048 | return _mm256_cmpneq_epi64_mask (_mm256_and_si256 (__A, __B), |
| 6049 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6050 | } |
| 6051 | |
| 6052 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6053 | _mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6054 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6055 | return _mm256_mask_cmpneq_epi64_mask (__U, _mm256_and_si256 (__A, __B), |
| 6056 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6057 | } |
| 6058 | |
| 6059 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6060 | _mm_testn_epi32_mask (__m128i __A, __m128i __B) |
| 6061 | { |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6062 | return _mm_cmpeq_epi32_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6063 | } |
| 6064 | |
| 6065 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6066 | _mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6067 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6068 | return _mm_mask_cmpeq_epi32_mask (__U, _mm_and_si128 (__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6069 | _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6070 | } |
| 6071 | |
| 6072 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6073 | _mm256_testn_epi32_mask (__m256i __A, __m256i __B) |
| 6074 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6075 | return _mm256_cmpeq_epi32_mask (_mm256_and_si256 (__A, __B), |
| 6076 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6077 | } |
| 6078 | |
| 6079 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6080 | _mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6081 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6082 | return _mm256_mask_cmpeq_epi32_mask (__U, _mm256_and_si256 (__A, __B), |
| 6083 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6084 | } |
| 6085 | |
| 6086 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6087 | _mm_testn_epi64_mask (__m128i __A, __m128i __B) |
| 6088 | { |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6089 | return _mm_cmpeq_epi64_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6090 | } |
| 6091 | |
| 6092 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6093 | _mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) |
| 6094 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6095 | return _mm_mask_cmpeq_epi64_mask (__U, _mm_and_si128 (__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6096 | _mm_setzero_si128()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6097 | } |
| 6098 | |
| 6099 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6100 | _mm256_testn_epi64_mask (__m256i __A, __m256i __B) |
| 6101 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6102 | return _mm256_cmpeq_epi64_mask (_mm256_and_si256 (__A, __B), |
| 6103 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6104 | } |
| 6105 | |
| 6106 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 6107 | _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) |
| 6108 | { |
Uriel Korach | 5b2b71d | 2017-11-13 12:50:52 +0000 | [diff] [blame] | 6109 | return _mm256_mask_cmpeq_epi64_mask (__U, _mm256_and_si256 (__A, __B), |
| 6110 | _mm256_setzero_si256()); |
Michael Zuckerman | 0752509 | 2016-04-11 10:22:07 +0000 | [diff] [blame] | 6111 | } |
| 6112 | |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6113 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6114 | _mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6115 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6116 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6117 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6118 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6119 | } |
| 6120 | |
| 6121 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6122 | _mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6123 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6124 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6125 | (__v4si)_mm_unpackhi_epi32(__A, __B), |
| 6126 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6127 | } |
| 6128 | |
| 6129 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6130 | _mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6131 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6132 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6133 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6134 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6135 | } |
| 6136 | |
| 6137 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6138 | _mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6139 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6140 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6141 | (__v8si)_mm256_unpackhi_epi32(__A, __B), |
| 6142 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6143 | } |
| 6144 | |
| 6145 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6146 | _mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6147 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6148 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6149 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
| 6150 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6151 | } |
| 6152 | |
| 6153 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6154 | _mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6155 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6156 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6157 | (__v2di)_mm_unpackhi_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6158 | (__v2di)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6159 | } |
| 6160 | |
| 6161 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6162 | _mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6163 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6164 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6165 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6166 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6167 | } |
| 6168 | |
| 6169 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6170 | _mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6171 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6172 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6173 | (__v4di)_mm256_unpackhi_epi64(__A, __B), |
| 6174 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6175 | } |
| 6176 | |
| 6177 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6178 | _mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6179 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6180 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6181 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6182 | (__v4si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6183 | } |
| 6184 | |
| 6185 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6186 | _mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6187 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6188 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6189 | (__v4si)_mm_unpacklo_epi32(__A, __B), |
| 6190 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6191 | } |
| 6192 | |
| 6193 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6194 | _mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6195 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6196 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6197 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6198 | (__v8si)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6199 | } |
| 6200 | |
| 6201 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6202 | _mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6203 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6204 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6205 | (__v8si)_mm256_unpacklo_epi32(__A, __B), |
| 6206 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6207 | } |
| 6208 | |
| 6209 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6210 | _mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6211 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6212 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6213 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
| 6214 | (__v2di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6215 | } |
| 6216 | |
| 6217 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6218 | _mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6219 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6220 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 6221 | (__v2di)_mm_unpacklo_epi64(__A, __B), |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6222 | (__v2di)_mm_setzero_si128()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6223 | } |
| 6224 | |
| 6225 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6226 | _mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6227 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6228 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6229 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6230 | (__v4di)__W); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6231 | } |
| 6232 | |
| 6233 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6234 | _mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B) |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6235 | { |
Craig Topper | 79f53ca | 2016-06-23 06:36:42 +0000 | [diff] [blame] | 6236 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 6237 | (__v4di)_mm256_unpacklo_epi64(__A, __B), |
| 6238 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 1af947a | 2016-04-11 12:32:31 +0000 | [diff] [blame] | 6239 | } |
| 6240 | |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6241 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6242 | _mm_mask_sra_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6243 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6244 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6245 | (__v4si)_mm_sra_epi32(__A, __B), |
| 6246 | (__v4si)__W); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6247 | } |
| 6248 | |
| 6249 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6250 | _mm_maskz_sra_epi32(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6251 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6252 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6253 | (__v4si)_mm_sra_epi32(__A, __B), |
| 6254 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6255 | } |
| 6256 | |
| 6257 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6258 | _mm256_mask_sra_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6259 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6260 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6261 | (__v8si)_mm256_sra_epi32(__A, __B), |
| 6262 | (__v8si)__W); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6263 | } |
| 6264 | |
| 6265 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6266 | _mm256_maskz_sra_epi32(__mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6267 | { |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6268 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6269 | (__v8si)_mm256_sra_epi32(__A, __B), |
| 6270 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6271 | } |
| 6272 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6273 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6274 | _mm_mask_srai_epi32(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
| 6275 | { |
| 6276 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6277 | (__v4si)_mm_srai_epi32(__A, __B), |
| 6278 | (__v4si)__W); |
| 6279 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6280 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6281 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6282 | _mm_maskz_srai_epi32(__mmask8 __U, __m128i __A, int __B) |
| 6283 | { |
| 6284 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 6285 | (__v4si)_mm_srai_epi32(__A, __B), |
| 6286 | (__v4si)_mm_setzero_si128()); |
| 6287 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6288 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6289 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6290 | _mm256_mask_srai_epi32(__m256i __W, __mmask8 __U, __m256i __A, int __B) |
| 6291 | { |
| 6292 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6293 | (__v8si)_mm256_srai_epi32(__A, __B), |
| 6294 | (__v8si)__W); |
| 6295 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6296 | |
Craig Topper | 66b2fd1 | 2016-10-31 04:30:51 +0000 | [diff] [blame] | 6297 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6298 | _mm256_maskz_srai_epi32(__mmask8 __U, __m256i __A, int __B) |
| 6299 | { |
| 6300 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 6301 | (__v8si)_mm256_srai_epi32(__A, __B), |
| 6302 | (__v8si)_mm256_setzero_si256()); |
| 6303 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6304 | |
| 6305 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6306 | _mm_sra_epi64(__m128i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6307 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6308 | return (__m128i)__builtin_ia32_psraq128((__v2di)__A, (__v2di)__B); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6309 | } |
| 6310 | |
| 6311 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6312 | _mm_mask_sra_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6313 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6314 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \ |
| 6315 | (__v2di)_mm_sra_epi64(__A, __B), \ |
| 6316 | (__v2di)__W); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6317 | } |
| 6318 | |
| 6319 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6320 | _mm_maskz_sra_epi64(__mmask8 __U, __m128i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6321 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6322 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \ |
| 6323 | (__v2di)_mm_sra_epi64(__A, __B), \ |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6324 | (__v2di)_mm_setzero_si128()); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6325 | } |
| 6326 | |
| 6327 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6328 | _mm256_sra_epi64(__m256i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6329 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6330 | return (__m256i)__builtin_ia32_psraq256((__v4di) __A, (__v2di) __B); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6331 | } |
| 6332 | |
| 6333 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6334 | _mm256_mask_sra_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6335 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6336 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \ |
| 6337 | (__v4di)_mm256_sra_epi64(__A, __B), \ |
| 6338 | (__v4di)__W); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6339 | } |
| 6340 | |
| 6341 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6342 | _mm256_maskz_sra_epi64(__mmask8 __U, __m256i __A, __m128i __B) |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6343 | { |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6344 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \ |
| 6345 | (__v4di)_mm256_sra_epi64(__A, __B), \ |
| 6346 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6347 | } |
| 6348 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6349 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6350 | _mm_srai_epi64(__m128i __A, int __imm) |
| 6351 | { |
| 6352 | return (__m128i)__builtin_ia32_psraqi128((__v2di)__A, __imm); |
| 6353 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6354 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6355 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6356 | _mm_mask_srai_epi64(__m128i __W, __mmask8 __U, __m128i __A, int __imm) |
| 6357 | { |
| 6358 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \ |
| 6359 | (__v2di)_mm_srai_epi64(__A, __imm), \ |
| 6360 | (__v2di)__W); |
| 6361 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6362 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6363 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6364 | _mm_maskz_srai_epi64(__mmask8 __U, __m128i __A, int __imm) |
| 6365 | { |
| 6366 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \ |
| 6367 | (__v2di)_mm_srai_epi64(__A, __imm), \ |
Craig Topper | dff5b31 | 2018-05-30 18:02:11 +0000 | [diff] [blame] | 6368 | (__v2di)_mm_setzero_si128()); |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6369 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6370 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6371 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6372 | _mm256_srai_epi64(__m256i __A, int __imm) |
| 6373 | { |
| 6374 | return (__m256i)__builtin_ia32_psraqi256((__v4di)__A, __imm); |
| 6375 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6376 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6377 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6378 | _mm256_mask_srai_epi64(__m256i __W, __mmask8 __U, __m256i __A, int __imm) |
| 6379 | { |
| 6380 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \ |
| 6381 | (__v4di)_mm256_srai_epi64(__A, __imm), \ |
| 6382 | (__v4di)__W); |
| 6383 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6384 | |
Craig Topper | 1a44193 | 2016-11-12 07:16:59 +0000 | [diff] [blame] | 6385 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6386 | _mm256_maskz_srai_epi64(__mmask8 __U, __m256i __A, int __imm) |
| 6387 | { |
| 6388 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \ |
| 6389 | (__v4di)_mm256_srai_epi64(__A, __imm), \ |
| 6390 | (__v4di)_mm256_setzero_si256()); |
| 6391 | } |
Michael Zuckerman | 6b5f4d8 | 2016-04-11 15:46:39 +0000 | [diff] [blame] | 6392 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6393 | #define _mm_ternarylogic_epi32(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6394 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 6395 | (__v4si)(__m128i)(B), \ |
| 6396 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6397 | (__mmask8)-1) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6398 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6399 | #define _mm_mask_ternarylogic_epi32(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6400 | (__m128i)__builtin_ia32_pternlogd128_mask((__v4si)(__m128i)(A), \ |
| 6401 | (__v4si)(__m128i)(B), \ |
| 6402 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6403 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6404 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6405 | #define _mm_maskz_ternarylogic_epi32(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6406 | (__m128i)__builtin_ia32_pternlogd128_maskz((__v4si)(__m128i)(A), \ |
| 6407 | (__v4si)(__m128i)(B), \ |
| 6408 | (__v4si)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6409 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6410 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6411 | #define _mm256_ternarylogic_epi32(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6412 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 6413 | (__v8si)(__m256i)(B), \ |
| 6414 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6415 | (__mmask8)-1) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6416 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6417 | #define _mm256_mask_ternarylogic_epi32(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6418 | (__m256i)__builtin_ia32_pternlogd256_mask((__v8si)(__m256i)(A), \ |
| 6419 | (__v8si)(__m256i)(B), \ |
| 6420 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6421 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6422 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6423 | #define _mm256_maskz_ternarylogic_epi32(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6424 | (__m256i)__builtin_ia32_pternlogd256_maskz((__v8si)(__m256i)(A), \ |
| 6425 | (__v8si)(__m256i)(B), \ |
| 6426 | (__v8si)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6427 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6428 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6429 | #define _mm_ternarylogic_epi64(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6430 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 6431 | (__v2di)(__m128i)(B), \ |
| 6432 | (__v2di)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6433 | (__mmask8)-1) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6434 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6435 | #define _mm_mask_ternarylogic_epi64(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6436 | (__m128i)__builtin_ia32_pternlogq128_mask((__v2di)(__m128i)(A), \ |
| 6437 | (__v2di)(__m128i)(B), \ |
| 6438 | (__v2di)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6439 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6440 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6441 | #define _mm_maskz_ternarylogic_epi64(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6442 | (__m128i)__builtin_ia32_pternlogq128_maskz((__v2di)(__m128i)(A), \ |
| 6443 | (__v2di)(__m128i)(B), \ |
| 6444 | (__v2di)(__m128i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6445 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6446 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6447 | #define _mm256_ternarylogic_epi64(A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6448 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 6449 | (__v4di)(__m256i)(B), \ |
| 6450 | (__v4di)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6451 | (__mmask8)-1) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6452 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6453 | #define _mm256_mask_ternarylogic_epi64(A, U, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6454 | (__m256i)__builtin_ia32_pternlogq256_mask((__v4di)(__m256i)(A), \ |
| 6455 | (__v4di)(__m256i)(B), \ |
| 6456 | (__v4di)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6457 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6458 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6459 | #define _mm256_maskz_ternarylogic_epi64(U, A, B, C, imm) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 6460 | (__m256i)__builtin_ia32_pternlogq256_maskz((__v4di)(__m256i)(A), \ |
| 6461 | (__v4di)(__m256i)(B), \ |
| 6462 | (__v4di)(__m256i)(C), (int)(imm), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6463 | (__mmask8)(U)) |
Michael Zuckerman | 81f468c | 2016-04-11 17:04:21 +0000 | [diff] [blame] | 6464 | |
| 6465 | |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6466 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6467 | #define _mm256_shuffle_f32x4(A, B, imm) \ |
Craig Topper | 9392136 | 2018-06-07 23:03:08 +0000 | [diff] [blame] | 6468 | (__m256)__builtin_ia32_shuf_f32x4_256((__v8sf)(__m256)(A), \ |
| 6469 | (__v8sf)(__m256)(B), (int)(imm)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6470 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6471 | #define _mm256_mask_shuffle_f32x4(W, U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6472 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6473 | (__v8sf)_mm256_shuffle_f32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6474 | (__v8sf)(__m256)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6475 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6476 | #define _mm256_maskz_shuffle_f32x4(U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6477 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6478 | (__v8sf)_mm256_shuffle_f32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6479 | (__v8sf)_mm256_setzero_ps()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6480 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6481 | #define _mm256_shuffle_f64x2(A, B, imm) \ |
Craig Topper | 9392136 | 2018-06-07 23:03:08 +0000 | [diff] [blame] | 6482 | (__m256d)__builtin_ia32_shuf_f64x2_256((__v4df)(__m256d)(A), \ |
| 6483 | (__v4df)(__m256d)(B), (int)(imm)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6484 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6485 | #define _mm256_mask_shuffle_f64x2(W, U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6486 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6487 | (__v4df)_mm256_shuffle_f64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6488 | (__v4df)(__m256)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6489 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6490 | #define _mm256_maskz_shuffle_f64x2(U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6491 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6492 | (__v4df)_mm256_shuffle_f64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6493 | (__v4df)_mm256_setzero_pd()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6494 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6495 | #define _mm256_shuffle_i32x4(A, B, imm) \ |
Craig Topper | 9392136 | 2018-06-07 23:03:08 +0000 | [diff] [blame] | 6496 | (__m256i)__builtin_ia32_shuf_i32x4_256((__v8si)(__m256i)(A), \ |
| 6497 | (__v8si)(__m256i)(B), (int)(imm)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6498 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6499 | #define _mm256_mask_shuffle_i32x4(W, U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6500 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 6501 | (__v8si)_mm256_shuffle_i32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6502 | (__v8si)(__m256)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6503 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6504 | #define _mm256_maskz_shuffle_i32x4(U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6505 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 6506 | (__v8si)_mm256_shuffle_i32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6507 | (__v8si)_mm256_setzero_si256()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6508 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6509 | #define _mm256_shuffle_i64x2(A, B, imm) \ |
Craig Topper | 9392136 | 2018-06-07 23:03:08 +0000 | [diff] [blame] | 6510 | (__m256i)__builtin_ia32_shuf_i64x2_256((__v4di)(__m256i)(A), \ |
| 6511 | (__v4di)(__m256i)(B), (int)(imm)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6512 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6513 | #define _mm256_mask_shuffle_i64x2(W, U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6514 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 6515 | (__v4di)_mm256_shuffle_i64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6516 | (__v4di)(__m256)(W)) |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6517 | |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6518 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6519 | #define _mm256_maskz_shuffle_i64x2(U, A, B, imm) \ |
Jina Nahias | dca9791 | 2017-11-13 09:15:31 +0000 | [diff] [blame] | 6520 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 6521 | (__v4di)_mm256_shuffle_i64x2((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6522 | (__v4di)_mm256_setzero_si256()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6523 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6524 | #define _mm_mask_shuffle_pd(W, U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6525 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6526 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6527 | (__v2df)(__m128d)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6528 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6529 | #define _mm_maskz_shuffle_pd(U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6530 | (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \ |
| 6531 | (__v2df)_mm_shuffle_pd((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6532 | (__v2df)_mm_setzero_pd()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6533 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6534 | #define _mm256_mask_shuffle_pd(W, U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6535 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6536 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6537 | (__v4df)(__m256d)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6538 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6539 | #define _mm256_maskz_shuffle_pd(U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6540 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 6541 | (__v4df)_mm256_shuffle_pd((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6542 | (__v4df)_mm256_setzero_pd()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6543 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6544 | #define _mm_mask_shuffle_ps(W, U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6545 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6546 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6547 | (__v4sf)(__m128)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6548 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6549 | #define _mm_maskz_shuffle_ps(U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6550 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 6551 | (__v4sf)_mm_shuffle_ps((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6552 | (__v4sf)_mm_setzero_ps()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6553 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6554 | #define _mm256_mask_shuffle_ps(W, U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6555 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6556 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6557 | (__v8sf)(__m256)(W)) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6558 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6559 | #define _mm256_maskz_shuffle_ps(U, A, B, M) \ |
Simon Pilgrim | 427154d | 2016-07-04 21:30:47 +0000 | [diff] [blame] | 6560 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 6561 | (__v8sf)_mm256_shuffle_ps((A), (B), (M)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 6562 | (__v8sf)_mm256_setzero_ps()) |
Michael Zuckerman | 04fb3bc | 2016-04-12 07:59:39 +0000 | [diff] [blame] | 6563 | |
| 6564 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6565 | _mm_rsqrt14_pd (__m128d __A) |
| 6566 | { |
| 6567 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 6568 | (__v2df) |
| 6569 | _mm_setzero_pd (), |
| 6570 | (__mmask8) -1); |
| 6571 | } |
| 6572 | |
| 6573 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6574 | _mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 6575 | { |
| 6576 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 6577 | (__v2df) __W, |
| 6578 | (__mmask8) __U); |
| 6579 | } |
| 6580 | |
| 6581 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 6582 | _mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A) |
| 6583 | { |
| 6584 | return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, |
| 6585 | (__v2df) |
| 6586 | _mm_setzero_pd (), |
| 6587 | (__mmask8) __U); |
| 6588 | } |
| 6589 | |
| 6590 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6591 | _mm256_rsqrt14_pd (__m256d __A) |
| 6592 | { |
| 6593 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 6594 | (__v4df) |
| 6595 | _mm256_setzero_pd (), |
| 6596 | (__mmask8) -1); |
| 6597 | } |
| 6598 | |
| 6599 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6600 | _mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 6601 | { |
| 6602 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 6603 | (__v4df) __W, |
| 6604 | (__mmask8) __U); |
| 6605 | } |
| 6606 | |
| 6607 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6608 | _mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A) |
| 6609 | { |
| 6610 | return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, |
| 6611 | (__v4df) |
| 6612 | _mm256_setzero_pd (), |
| 6613 | (__mmask8) __U); |
| 6614 | } |
| 6615 | |
| 6616 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6617 | _mm_rsqrt14_ps (__m128 __A) |
| 6618 | { |
| 6619 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 6620 | (__v4sf) |
| 6621 | _mm_setzero_ps (), |
| 6622 | (__mmask8) -1); |
| 6623 | } |
| 6624 | |
| 6625 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6626 | _mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 6627 | { |
| 6628 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 6629 | (__v4sf) __W, |
| 6630 | (__mmask8) __U); |
| 6631 | } |
| 6632 | |
| 6633 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6634 | _mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A) |
| 6635 | { |
| 6636 | return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, |
| 6637 | (__v4sf) |
| 6638 | _mm_setzero_ps (), |
| 6639 | (__mmask8) __U); |
| 6640 | } |
| 6641 | |
| 6642 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6643 | _mm256_rsqrt14_ps (__m256 __A) |
| 6644 | { |
| 6645 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 6646 | (__v8sf) |
| 6647 | _mm256_setzero_ps (), |
| 6648 | (__mmask8) -1); |
| 6649 | } |
| 6650 | |
| 6651 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6652 | _mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 6653 | { |
| 6654 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 6655 | (__v8sf) __W, |
| 6656 | (__mmask8) __U); |
| 6657 | } |
| 6658 | |
| 6659 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6660 | _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A) |
| 6661 | { |
| 6662 | return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, |
| 6663 | (__v8sf) |
| 6664 | _mm256_setzero_ps (), |
| 6665 | (__mmask8) __U); |
| 6666 | } |
| 6667 | |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6668 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6669 | _mm256_broadcast_f32x4(__m128 __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6670 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6671 | return (__m256)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A, |
| 6672 | 0, 1, 2, 3, 0, 1, 2, 3); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6673 | } |
| 6674 | |
| 6675 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6676 | _mm256_mask_broadcast_f32x4(__m256 __O, __mmask8 __M, __m128 __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6677 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6678 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 6679 | (__v8sf)_mm256_broadcast_f32x4(__A), |
| 6680 | (__v8sf)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6681 | } |
| 6682 | |
| 6683 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6684 | _mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A) |
| 6685 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6686 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 6687 | (__v8sf)_mm256_broadcast_f32x4(__A), |
| 6688 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6689 | } |
| 6690 | |
| 6691 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6692 | _mm256_broadcast_i32x4(__m128i __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6693 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6694 | return (__m256i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
| 6695 | 0, 1, 2, 3, 0, 1, 2, 3); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6696 | } |
| 6697 | |
| 6698 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6699 | _mm256_mask_broadcast_i32x4(__m256i __O, __mmask8 __M, __m128i __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6700 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6701 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 6702 | (__v8si)_mm256_broadcast_i32x4(__A), |
| 6703 | (__v8si)__O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6704 | } |
| 6705 | |
| 6706 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6707 | _mm256_maskz_broadcast_i32x4(__mmask8 __M, __m128i __A) |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6708 | { |
Craig Topper | 367c86d | 2017-01-18 02:17:10 +0000 | [diff] [blame] | 6709 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 6710 | (__v8si)_mm256_broadcast_i32x4(__A), |
| 6711 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6712 | } |
| 6713 | |
| 6714 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6715 | _mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A) |
| 6716 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6717 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 6718 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 6719 | (__v4df) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6720 | } |
| 6721 | |
| 6722 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 6723 | _mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A) |
| 6724 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6725 | return (__m256d)__builtin_ia32_selectpd_256(__M, |
| 6726 | (__v4df) _mm256_broadcastsd_pd(__A), |
| 6727 | (__v4df) _mm256_setzero_pd()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6728 | } |
| 6729 | |
| 6730 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6731 | _mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A) |
| 6732 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6733 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 6734 | (__v4sf) _mm_broadcastss_ps(__A), |
| 6735 | (__v4sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6736 | } |
| 6737 | |
| 6738 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 6739 | _mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 6740 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6741 | return (__m128)__builtin_ia32_selectps_128(__M, |
| 6742 | (__v4sf) _mm_broadcastss_ps(__A), |
| 6743 | (__v4sf) _mm_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6744 | } |
| 6745 | |
| 6746 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6747 | _mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A) |
| 6748 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6749 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 6750 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 6751 | (__v8sf) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6752 | } |
| 6753 | |
| 6754 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 6755 | _mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) |
| 6756 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6757 | return (__m256)__builtin_ia32_selectps_256(__M, |
| 6758 | (__v8sf) _mm256_broadcastss_ps(__A), |
| 6759 | (__v8sf) _mm256_setzero_ps()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6760 | } |
| 6761 | |
| 6762 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6763 | _mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 6764 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6765 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 6766 | (__v4si) _mm_broadcastd_epi32(__A), |
| 6767 | (__v4si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6768 | } |
| 6769 | |
| 6770 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6771 | _mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 6772 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6773 | return (__m128i)__builtin_ia32_selectd_128(__M, |
| 6774 | (__v4si) _mm_broadcastd_epi32(__A), |
| 6775 | (__v4si) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6776 | } |
| 6777 | |
| 6778 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6779 | _mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A) |
| 6780 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6781 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 6782 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 6783 | (__v8si) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6784 | } |
| 6785 | |
| 6786 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6787 | _mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) |
| 6788 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6789 | return (__m256i)__builtin_ia32_selectd_256(__M, |
| 6790 | (__v8si) _mm256_broadcastd_epi32(__A), |
| 6791 | (__v8si) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6792 | } |
| 6793 | |
| 6794 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6795 | _mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A) |
| 6796 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6797 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 6798 | (__v2di) _mm_broadcastq_epi64(__A), |
| 6799 | (__v2di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6800 | } |
| 6801 | |
| 6802 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6803 | _mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 6804 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6805 | return (__m128i)__builtin_ia32_selectq_128(__M, |
| 6806 | (__v2di) _mm_broadcastq_epi64(__A), |
| 6807 | (__v2di) _mm_setzero_si128()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6808 | } |
| 6809 | |
| 6810 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6811 | _mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A) |
| 6812 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6813 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 6814 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 6815 | (__v4di) __O); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6816 | } |
| 6817 | |
| 6818 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 6819 | _mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) |
| 6820 | { |
Simon Pilgrim | f5a8837 | 2016-07-05 12:59:33 +0000 | [diff] [blame] | 6821 | return (__m256i)__builtin_ia32_selectq_256(__M, |
| 6822 | (__v4di) _mm256_broadcastq_epi64(__A), |
| 6823 | (__v4di) _mm256_setzero_si256()); |
Michael Zuckerman | c2b6128 | 2016-04-13 12:58:01 +0000 | [diff] [blame] | 6824 | } |
| 6825 | |
Michael Zuckerman | e168061 | 2016-04-13 15:02:04 +0000 | [diff] [blame] | 6826 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6827 | _mm_cvtsepi32_epi8 (__m128i __A) |
| 6828 | { |
| 6829 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 6830 | (__v16qi)_mm_undefined_si128(), |
| 6831 | (__mmask8) -1); |
| 6832 | } |
| 6833 | |
| 6834 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6835 | _mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 6836 | { |
| 6837 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 6838 | (__v16qi) __O, __M); |
| 6839 | } |
| 6840 | |
| 6841 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6842 | _mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A) |
| 6843 | { |
| 6844 | return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, |
| 6845 | (__v16qi) _mm_setzero_si128 (), |
| 6846 | __M); |
| 6847 | } |
| 6848 | |
| 6849 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6850 | _mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 6851 | { |
| 6852 | __builtin_ia32_pmovsdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 6853 | } |
| 6854 | |
| 6855 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6856 | _mm256_cvtsepi32_epi8 (__m256i __A) |
| 6857 | { |
| 6858 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 6859 | (__v16qi)_mm_undefined_si128(), |
| 6860 | (__mmask8) -1); |
| 6861 | } |
| 6862 | |
| 6863 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6864 | _mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 6865 | { |
| 6866 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 6867 | (__v16qi) __O, __M); |
| 6868 | } |
| 6869 | |
| 6870 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6871 | _mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A) |
| 6872 | { |
| 6873 | return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, |
| 6874 | (__v16qi) _mm_setzero_si128 (), |
| 6875 | __M); |
| 6876 | } |
| 6877 | |
| 6878 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6879 | _mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 6880 | { |
| 6881 | __builtin_ia32_pmovsdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 6882 | } |
| 6883 | |
| 6884 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6885 | _mm_cvtsepi32_epi16 (__m128i __A) |
| 6886 | { |
| 6887 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 6888 | (__v8hi)_mm_setzero_si128 (), |
| 6889 | (__mmask8) -1); |
| 6890 | } |
| 6891 | |
| 6892 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6893 | _mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 6894 | { |
| 6895 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 6896 | (__v8hi)__O, |
| 6897 | __M); |
| 6898 | } |
| 6899 | |
| 6900 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6901 | _mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A) |
| 6902 | { |
| 6903 | return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, |
| 6904 | (__v8hi) _mm_setzero_si128 (), |
| 6905 | __M); |
| 6906 | } |
| 6907 | |
| 6908 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6909 | _mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 6910 | { |
| 6911 | __builtin_ia32_pmovsdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 6912 | } |
| 6913 | |
| 6914 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6915 | _mm256_cvtsepi32_epi16 (__m256i __A) |
| 6916 | { |
| 6917 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 6918 | (__v8hi)_mm_undefined_si128(), |
| 6919 | (__mmask8) -1); |
| 6920 | } |
| 6921 | |
| 6922 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6923 | _mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 6924 | { |
| 6925 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 6926 | (__v8hi) __O, __M); |
| 6927 | } |
| 6928 | |
| 6929 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6930 | _mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A) |
| 6931 | { |
| 6932 | return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, |
| 6933 | (__v8hi) _mm_setzero_si128 (), |
| 6934 | __M); |
| 6935 | } |
| 6936 | |
| 6937 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6938 | _mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 6939 | { |
| 6940 | __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 6941 | } |
| 6942 | |
| 6943 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6944 | _mm_cvtsepi64_epi8 (__m128i __A) |
| 6945 | { |
| 6946 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 6947 | (__v16qi)_mm_undefined_si128(), |
| 6948 | (__mmask8) -1); |
| 6949 | } |
| 6950 | |
| 6951 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6952 | _mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 6953 | { |
| 6954 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 6955 | (__v16qi) __O, __M); |
| 6956 | } |
| 6957 | |
| 6958 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6959 | _mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A) |
| 6960 | { |
| 6961 | return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, |
| 6962 | (__v16qi) _mm_setzero_si128 (), |
| 6963 | __M); |
| 6964 | } |
| 6965 | |
| 6966 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6967 | _mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 6968 | { |
| 6969 | __builtin_ia32_pmovsqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 6970 | } |
| 6971 | |
| 6972 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6973 | _mm256_cvtsepi64_epi8 (__m256i __A) |
| 6974 | { |
| 6975 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 6976 | (__v16qi)_mm_undefined_si128(), |
| 6977 | (__mmask8) -1); |
| 6978 | } |
| 6979 | |
| 6980 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6981 | _mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 6982 | { |
| 6983 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 6984 | (__v16qi) __O, __M); |
| 6985 | } |
| 6986 | |
| 6987 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 6988 | _mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A) |
| 6989 | { |
| 6990 | return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, |
| 6991 | (__v16qi) _mm_setzero_si128 (), |
| 6992 | __M); |
| 6993 | } |
| 6994 | |
| 6995 | static __inline__ void __DEFAULT_FN_ATTRS |
| 6996 | _mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 6997 | { |
| 6998 | __builtin_ia32_pmovsqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 6999 | } |
| 7000 | |
| 7001 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7002 | _mm_cvtsepi64_epi32 (__m128i __A) |
| 7003 | { |
| 7004 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7005 | (__v4si)_mm_undefined_si128(), |
| 7006 | (__mmask8) -1); |
| 7007 | } |
| 7008 | |
| 7009 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7010 | _mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7011 | { |
| 7012 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7013 | (__v4si) __O, __M); |
| 7014 | } |
| 7015 | |
| 7016 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7017 | _mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7018 | { |
| 7019 | return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, |
| 7020 | (__v4si) _mm_setzero_si128 (), |
| 7021 | __M); |
| 7022 | } |
| 7023 | |
| 7024 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7025 | _mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 7026 | { |
| 7027 | __builtin_ia32_pmovsqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 7028 | } |
| 7029 | |
| 7030 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7031 | _mm256_cvtsepi64_epi32 (__m256i __A) |
| 7032 | { |
| 7033 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7034 | (__v4si)_mm_undefined_si128(), |
| 7035 | (__mmask8) -1); |
| 7036 | } |
| 7037 | |
| 7038 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7039 | _mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7040 | { |
| 7041 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7042 | (__v4si)__O, |
| 7043 | __M); |
| 7044 | } |
| 7045 | |
| 7046 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7047 | _mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A) |
| 7048 | { |
| 7049 | return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, |
| 7050 | (__v4si) _mm_setzero_si128 (), |
| 7051 | __M); |
| 7052 | } |
| 7053 | |
| 7054 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7055 | _mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 7056 | { |
| 7057 | __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 7058 | } |
| 7059 | |
| 7060 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7061 | _mm_cvtsepi64_epi16 (__m128i __A) |
| 7062 | { |
| 7063 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7064 | (__v8hi)_mm_undefined_si128(), |
| 7065 | (__mmask8) -1); |
| 7066 | } |
| 7067 | |
| 7068 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7069 | _mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7070 | { |
| 7071 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7072 | (__v8hi) __O, __M); |
| 7073 | } |
| 7074 | |
| 7075 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7076 | _mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A) |
| 7077 | { |
| 7078 | return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, |
| 7079 | (__v8hi) _mm_setzero_si128 (), |
| 7080 | __M); |
| 7081 | } |
| 7082 | |
| 7083 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7084 | _mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7085 | { |
| 7086 | __builtin_ia32_pmovsqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 7087 | } |
| 7088 | |
| 7089 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7090 | _mm256_cvtsepi64_epi16 (__m256i __A) |
| 7091 | { |
| 7092 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7093 | (__v8hi)_mm_undefined_si128(), |
| 7094 | (__mmask8) -1); |
| 7095 | } |
| 7096 | |
| 7097 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7098 | _mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7099 | { |
| 7100 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7101 | (__v8hi) __O, __M); |
| 7102 | } |
| 7103 | |
| 7104 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7105 | _mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A) |
| 7106 | { |
| 7107 | return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, |
| 7108 | (__v8hi) _mm_setzero_si128 (), |
| 7109 | __M); |
| 7110 | } |
| 7111 | |
| 7112 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7113 | _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7114 | { |
| 7115 | __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 7116 | } |
| 7117 | |
Michael Zuckerman | d871531 | 2016-04-14 06:48:09 +0000 | [diff] [blame] | 7118 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7119 | _mm_cvtusepi32_epi8 (__m128i __A) |
| 7120 | { |
| 7121 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7122 | (__v16qi)_mm_undefined_si128(), |
| 7123 | (__mmask8) -1); |
| 7124 | } |
| 7125 | |
| 7126 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7127 | _mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7128 | { |
| 7129 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7130 | (__v16qi) __O, |
| 7131 | __M); |
| 7132 | } |
| 7133 | |
| 7134 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7135 | _mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7136 | { |
| 7137 | return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, |
| 7138 | (__v16qi) _mm_setzero_si128 (), |
| 7139 | __M); |
| 7140 | } |
| 7141 | |
| 7142 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7143 | _mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7144 | { |
| 7145 | __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7146 | } |
| 7147 | |
| 7148 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7149 | _mm256_cvtusepi32_epi8 (__m256i __A) |
| 7150 | { |
| 7151 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7152 | (__v16qi)_mm_undefined_si128(), |
| 7153 | (__mmask8) -1); |
| 7154 | } |
| 7155 | |
| 7156 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7157 | _mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7158 | { |
| 7159 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7160 | (__v16qi) __O, |
| 7161 | __M); |
| 7162 | } |
| 7163 | |
| 7164 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7165 | _mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7166 | { |
| 7167 | return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, |
| 7168 | (__v16qi) _mm_setzero_si128 (), |
| 7169 | __M); |
| 7170 | } |
| 7171 | |
| 7172 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7173 | _mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7174 | { |
| 7175 | __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M); |
| 7176 | } |
| 7177 | |
| 7178 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7179 | _mm_cvtusepi32_epi16 (__m128i __A) |
| 7180 | { |
| 7181 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7182 | (__v8hi)_mm_undefined_si128(), |
| 7183 | (__mmask8) -1); |
| 7184 | } |
| 7185 | |
| 7186 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7187 | _mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7188 | { |
| 7189 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7190 | (__v8hi) __O, __M); |
| 7191 | } |
| 7192 | |
| 7193 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7194 | _mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7195 | { |
| 7196 | return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, |
| 7197 | (__v8hi) _mm_setzero_si128 (), |
| 7198 | __M); |
| 7199 | } |
| 7200 | |
| 7201 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7202 | _mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7203 | { |
| 7204 | __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7205 | } |
| 7206 | |
| 7207 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7208 | _mm256_cvtusepi32_epi16 (__m256i __A) |
| 7209 | { |
| 7210 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7211 | (__v8hi) _mm_undefined_si128(), |
| 7212 | (__mmask8) -1); |
| 7213 | } |
| 7214 | |
| 7215 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7216 | _mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7217 | { |
| 7218 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7219 | (__v8hi) __O, __M); |
| 7220 | } |
| 7221 | |
| 7222 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7223 | _mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7224 | { |
| 7225 | return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, |
| 7226 | (__v8hi) _mm_setzero_si128 (), |
| 7227 | __M); |
| 7228 | } |
| 7229 | |
| 7230 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7231 | _mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7232 | { |
| 7233 | __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 7234 | } |
| 7235 | |
| 7236 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7237 | _mm_cvtusepi64_epi8 (__m128i __A) |
| 7238 | { |
| 7239 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7240 | (__v16qi)_mm_undefined_si128(), |
| 7241 | (__mmask8) -1); |
| 7242 | } |
| 7243 | |
| 7244 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7245 | _mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7246 | { |
| 7247 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7248 | (__v16qi) __O, |
| 7249 | __M); |
| 7250 | } |
| 7251 | |
| 7252 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7253 | _mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A) |
| 7254 | { |
| 7255 | return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, |
| 7256 | (__v16qi) _mm_setzero_si128 (), |
| 7257 | __M); |
| 7258 | } |
| 7259 | |
| 7260 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7261 | _mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7262 | { |
| 7263 | __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 7264 | } |
| 7265 | |
| 7266 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7267 | _mm256_cvtusepi64_epi8 (__m256i __A) |
| 7268 | { |
| 7269 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7270 | (__v16qi)_mm_undefined_si128(), |
| 7271 | (__mmask8) -1); |
| 7272 | } |
| 7273 | |
| 7274 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7275 | _mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7276 | { |
| 7277 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7278 | (__v16qi) __O, |
| 7279 | __M); |
| 7280 | } |
| 7281 | |
| 7282 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7283 | _mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A) |
| 7284 | { |
| 7285 | return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, |
| 7286 | (__v16qi) _mm_setzero_si128 (), |
| 7287 | __M); |
| 7288 | } |
| 7289 | |
| 7290 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7291 | _mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7292 | { |
| 7293 | __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 7294 | } |
| 7295 | |
| 7296 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7297 | _mm_cvtusepi64_epi32 (__m128i __A) |
| 7298 | { |
| 7299 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7300 | (__v4si)_mm_undefined_si128(), |
| 7301 | (__mmask8) -1); |
| 7302 | } |
| 7303 | |
| 7304 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7305 | _mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7306 | { |
| 7307 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7308 | (__v4si) __O, __M); |
| 7309 | } |
| 7310 | |
| 7311 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7312 | _mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7313 | { |
| 7314 | return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, |
| 7315 | (__v4si) _mm_setzero_si128 (), |
| 7316 | __M); |
| 7317 | } |
| 7318 | |
| 7319 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7320 | _mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 7321 | { |
| 7322 | __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 7323 | } |
| 7324 | |
| 7325 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7326 | _mm256_cvtusepi64_epi32 (__m256i __A) |
| 7327 | { |
| 7328 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 7329 | (__v4si)_mm_undefined_si128(), |
| 7330 | (__mmask8) -1); |
| 7331 | } |
| 7332 | |
| 7333 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7334 | _mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7335 | { |
| 7336 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 7337 | (__v4si) __O, __M); |
| 7338 | } |
| 7339 | |
| 7340 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7341 | _mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A) |
| 7342 | { |
| 7343 | return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, |
| 7344 | (__v4si) _mm_setzero_si128 (), |
| 7345 | __M); |
| 7346 | } |
| 7347 | |
| 7348 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7349 | _mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 7350 | { |
| 7351 | __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 7352 | } |
| 7353 | |
| 7354 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7355 | _mm_cvtusepi64_epi16 (__m128i __A) |
| 7356 | { |
| 7357 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 7358 | (__v8hi)_mm_undefined_si128(), |
| 7359 | (__mmask8) -1); |
| 7360 | } |
| 7361 | |
| 7362 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7363 | _mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7364 | { |
| 7365 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 7366 | (__v8hi) __O, __M); |
| 7367 | } |
| 7368 | |
| 7369 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7370 | _mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A) |
| 7371 | { |
| 7372 | return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, |
| 7373 | (__v8hi) _mm_setzero_si128 (), |
| 7374 | __M); |
| 7375 | } |
| 7376 | |
| 7377 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7378 | _mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7379 | { |
| 7380 | __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 7381 | } |
| 7382 | |
| 7383 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7384 | _mm256_cvtusepi64_epi16 (__m256i __A) |
| 7385 | { |
| 7386 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 7387 | (__v8hi)_mm_undefined_si128(), |
| 7388 | (__mmask8) -1); |
| 7389 | } |
| 7390 | |
| 7391 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7392 | _mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7393 | { |
| 7394 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 7395 | (__v8hi) __O, __M); |
| 7396 | } |
| 7397 | |
| 7398 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7399 | _mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A) |
| 7400 | { |
| 7401 | return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, |
| 7402 | (__v8hi) _mm_setzero_si128 (), |
| 7403 | __M); |
| 7404 | } |
| 7405 | |
| 7406 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7407 | _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7408 | { |
Craig Topper | 819f2a2 | 2018-05-30 17:23:45 +0000 | [diff] [blame] | 7409 | __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
Michael Zuckerman | d871531 | 2016-04-14 06:48:09 +0000 | [diff] [blame] | 7410 | } |
| 7411 | |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7412 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7413 | _mm_cvtepi32_epi8 (__m128i __A) |
| 7414 | { |
| 7415 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 7416 | (__v16qi)_mm_undefined_si128(), |
| 7417 | (__mmask8) -1); |
| 7418 | } |
| 7419 | |
| 7420 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7421 | _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7422 | { |
| 7423 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 7424 | (__v16qi) __O, __M); |
| 7425 | } |
| 7426 | |
| 7427 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7428 | _mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A) |
| 7429 | { |
| 7430 | return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, |
| 7431 | (__v16qi) |
| 7432 | _mm_setzero_si128 (), |
| 7433 | __M); |
| 7434 | } |
| 7435 | |
| 7436 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7437 | _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7438 | { |
| 7439 | __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); |
| 7440 | } |
| 7441 | |
| 7442 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7443 | _mm256_cvtepi32_epi8 (__m256i __A) |
| 7444 | { |
| 7445 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 7446 | (__v16qi)_mm_undefined_si128(), |
| 7447 | (__mmask8) -1); |
| 7448 | } |
| 7449 | |
| 7450 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7451 | _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7452 | { |
| 7453 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 7454 | (__v16qi) __O, __M); |
| 7455 | } |
| 7456 | |
| 7457 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7458 | _mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A) |
| 7459 | { |
| 7460 | return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, |
| 7461 | (__v16qi) _mm_setzero_si128 (), |
| 7462 | __M); |
| 7463 | } |
| 7464 | |
| 7465 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7466 | _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7467 | { |
| 7468 | __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M); |
| 7469 | } |
| 7470 | |
| 7471 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7472 | _mm_cvtepi32_epi16 (__m128i __A) |
| 7473 | { |
| 7474 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 7475 | (__v8hi) _mm_setzero_si128 (), |
| 7476 | (__mmask8) -1); |
| 7477 | } |
| 7478 | |
| 7479 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7480 | _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7481 | { |
| 7482 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 7483 | (__v8hi) __O, __M); |
| 7484 | } |
| 7485 | |
| 7486 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7487 | _mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A) |
| 7488 | { |
| 7489 | return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, |
| 7490 | (__v8hi) _mm_setzero_si128 (), |
| 7491 | __M); |
| 7492 | } |
| 7493 | |
| 7494 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7495 | _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7496 | { |
| 7497 | __builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); |
| 7498 | } |
| 7499 | |
| 7500 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7501 | _mm256_cvtepi32_epi16 (__m256i __A) |
| 7502 | { |
Craig Topper | 25de41c | 2018-05-14 17:50:40 +0000 | [diff] [blame] | 7503 | return (__m128i)__builtin_convertvector((__v8si)__A, __v8hi); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7504 | } |
| 7505 | |
| 7506 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7507 | _mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7508 | { |
Craig Topper | 9d146bb | 2018-05-15 03:17:52 +0000 | [diff] [blame] | 7509 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 7510 | (__v8hi) __O, __M); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7511 | } |
| 7512 | |
| 7513 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7514 | _mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A) |
| 7515 | { |
Craig Topper | 9d146bb | 2018-05-15 03:17:52 +0000 | [diff] [blame] | 7516 | return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, |
| 7517 | (__v8hi) _mm_setzero_si128 (), |
| 7518 | __M); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7519 | } |
| 7520 | |
| 7521 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7522 | _mm256_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7523 | { |
| 7524 | __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); |
| 7525 | } |
| 7526 | |
| 7527 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7528 | _mm_cvtepi64_epi8 (__m128i __A) |
| 7529 | { |
| 7530 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 7531 | (__v16qi) _mm_undefined_si128(), |
| 7532 | (__mmask8) -1); |
| 7533 | } |
| 7534 | |
| 7535 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7536 | _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7537 | { |
| 7538 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 7539 | (__v16qi) __O, __M); |
| 7540 | } |
| 7541 | |
| 7542 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7543 | _mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A) |
| 7544 | { |
| 7545 | return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, |
| 7546 | (__v16qi) _mm_setzero_si128 (), |
| 7547 | __M); |
| 7548 | } |
| 7549 | |
| 7550 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7551 | _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
| 7552 | { |
| 7553 | __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); |
| 7554 | } |
| 7555 | |
| 7556 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7557 | _mm256_cvtepi64_epi8 (__m256i __A) |
| 7558 | { |
| 7559 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 7560 | (__v16qi) _mm_undefined_si128(), |
| 7561 | (__mmask8) -1); |
| 7562 | } |
| 7563 | |
| 7564 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7565 | _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7566 | { |
| 7567 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 7568 | (__v16qi) __O, __M); |
| 7569 | } |
| 7570 | |
| 7571 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7572 | _mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A) |
| 7573 | { |
| 7574 | return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, |
| 7575 | (__v16qi) _mm_setzero_si128 (), |
| 7576 | __M); |
| 7577 | } |
| 7578 | |
| 7579 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7580 | _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) |
| 7581 | { |
| 7582 | __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); |
| 7583 | } |
| 7584 | |
| 7585 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7586 | _mm_cvtepi64_epi32 (__m128i __A) |
| 7587 | { |
| 7588 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 7589 | (__v4si)_mm_undefined_si128(), |
| 7590 | (__mmask8) -1); |
| 7591 | } |
| 7592 | |
| 7593 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7594 | _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7595 | { |
| 7596 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 7597 | (__v4si) __O, __M); |
| 7598 | } |
| 7599 | |
| 7600 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7601 | _mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A) |
| 7602 | { |
| 7603 | return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, |
| 7604 | (__v4si) _mm_setzero_si128 (), |
| 7605 | __M); |
| 7606 | } |
| 7607 | |
| 7608 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7609 | _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) |
| 7610 | { |
| 7611 | __builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); |
| 7612 | } |
| 7613 | |
| 7614 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7615 | _mm256_cvtepi64_epi32 (__m256i __A) |
| 7616 | { |
Craig Topper | 25de41c | 2018-05-14 17:50:40 +0000 | [diff] [blame] | 7617 | return (__m128i)__builtin_convertvector((__v4di)__A, __v4si); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7618 | } |
| 7619 | |
| 7620 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7621 | _mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7622 | { |
Craig Topper | 25de41c | 2018-05-14 17:50:40 +0000 | [diff] [blame] | 7623 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 7624 | (__v4si)_mm256_cvtepi64_epi32(__A), |
| 7625 | (__v4si)__O); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7626 | } |
| 7627 | |
| 7628 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7629 | _mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A) |
| 7630 | { |
Craig Topper | 25de41c | 2018-05-14 17:50:40 +0000 | [diff] [blame] | 7631 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 7632 | (__v4si)_mm256_cvtepi64_epi32(__A), |
| 7633 | (__v4si)_mm_setzero_si128()); |
Michael Zuckerman | 0a3508a | 2016-04-14 07:56:51 +0000 | [diff] [blame] | 7634 | } |
| 7635 | |
| 7636 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7637 | _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) |
| 7638 | { |
| 7639 | __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); |
| 7640 | } |
| 7641 | |
| 7642 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7643 | _mm_cvtepi64_epi16 (__m128i __A) |
| 7644 | { |
| 7645 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 7646 | (__v8hi) _mm_undefined_si128(), |
| 7647 | (__mmask8) -1); |
| 7648 | } |
| 7649 | |
| 7650 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7651 | _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
| 7652 | { |
| 7653 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 7654 | (__v8hi)__O, |
| 7655 | __M); |
| 7656 | } |
| 7657 | |
| 7658 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7659 | _mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A) |
| 7660 | { |
| 7661 | return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, |
| 7662 | (__v8hi) _mm_setzero_si128 (), |
| 7663 | __M); |
| 7664 | } |
| 7665 | |
| 7666 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7667 | _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) |
| 7668 | { |
| 7669 | __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); |
| 7670 | } |
| 7671 | |
| 7672 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7673 | _mm256_cvtepi64_epi16 (__m256i __A) |
| 7674 | { |
| 7675 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 7676 | (__v8hi)_mm_undefined_si128(), |
| 7677 | (__mmask8) -1); |
| 7678 | } |
| 7679 | |
| 7680 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7681 | _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) |
| 7682 | { |
| 7683 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 7684 | (__v8hi) __O, __M); |
| 7685 | } |
| 7686 | |
| 7687 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
| 7688 | _mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A) |
| 7689 | { |
| 7690 | return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, |
| 7691 | (__v8hi) _mm_setzero_si128 (), |
| 7692 | __M); |
| 7693 | } |
| 7694 | |
| 7695 | static __inline__ void __DEFAULT_FN_ATTRS |
| 7696 | _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) |
| 7697 | { |
| 7698 | __builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); |
| 7699 | } |
| 7700 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7701 | #define _mm256_extractf32x4_ps(A, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame^] | 7702 | (__m128)__builtin_ia32_extractf32x4_256((__v8sf)(__m256)(A), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7703 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7704 | #define _mm256_mask_extractf32x4_ps(W, U, A, imm) \ |
Craig Topper | 93ffabd | 2016-10-31 04:30:56 +0000 | [diff] [blame] | 7705 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7706 | (__v4sf)_mm256_extractf32x4_ps((A), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 7707 | (__v4sf)(__m128)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7708 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7709 | #define _mm256_maskz_extractf32x4_ps(U, A, imm) \ |
Craig Topper | 93ffabd | 2016-10-31 04:30:56 +0000 | [diff] [blame] | 7710 | (__m128)__builtin_ia32_selectps_128((__mmask8)(U), \ |
| 7711 | (__v4sf)_mm256_extractf32x4_ps((A), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7712 | (__v4sf)_mm_setzero_ps()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7713 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7714 | #define _mm256_extracti32x4_epi32(A, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame^] | 7715 | (__m128i)__builtin_ia32_extracti32x4_256((__v8si)(__m256i)(A), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7716 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7717 | #define _mm256_mask_extracti32x4_epi32(W, U, A, imm) \ |
Craig Topper | 3507296 | 2016-10-31 05:49:11 +0000 | [diff] [blame] | 7718 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
Craig Topper | 93ffabd | 2016-10-31 04:30:56 +0000 | [diff] [blame] | 7719 | (__v4si)_mm256_extracti32x4_epi32((A), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 7720 | (__v4si)(__m128i)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7721 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7722 | #define _mm256_maskz_extracti32x4_epi32(U, A, imm) \ |
Craig Topper | 3507296 | 2016-10-31 05:49:11 +0000 | [diff] [blame] | 7723 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
Craig Topper | 93ffabd | 2016-10-31 04:30:56 +0000 | [diff] [blame] | 7724 | (__v4si)_mm256_extracti32x4_epi32((A), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7725 | (__v4si)_mm_setzero_si128()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7726 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7727 | #define _mm256_insertf32x4(A, B, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame^] | 7728 | (__m256)__builtin_ia32_insertf32x4_256((__v8sf)(__m256)(A), \ |
| 7729 | (__v4sf)(__m128)(B), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7730 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7731 | #define _mm256_mask_insertf32x4(W, U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 7732 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7733 | (__v8sf)_mm256_insertf32x4((A), (B), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 7734 | (__v8sf)(__m256)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7735 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7736 | #define _mm256_maskz_insertf32x4(U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 7737 | (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \ |
| 7738 | (__v8sf)_mm256_insertf32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7739 | (__v8sf)_mm256_setzero_ps()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7740 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7741 | #define _mm256_inserti32x4(A, B, imm) \ |
Craig Topper | 3428bee | 2018-06-08 03:24:47 +0000 | [diff] [blame^] | 7742 | (__m256i)__builtin_ia32_inserti32x4_256((__v8si)(__m256i)(A), \ |
| 7743 | (__v4si)(__m128i)(B), (int)(imm)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7744 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7745 | #define _mm256_mask_inserti32x4(W, U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 7746 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 7747 | (__v8si)_mm256_inserti32x4((A), (B), (imm)), \ |
Craig Topper | cbf3929 | 2018-05-31 01:24:40 +0000 | [diff] [blame] | 7748 | (__v8si)(__m256i)(W)) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7749 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7750 | #define _mm256_maskz_inserti32x4(U, A, B, imm) \ |
Craig Topper | 08bf53f | 2016-11-01 05:47:56 +0000 | [diff] [blame] | 7751 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 7752 | (__v8si)_mm256_inserti32x4((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7753 | (__v8si)_mm256_setzero_si256()) |
Michael Zuckerman | ef2979a | 2016-04-19 15:18:23 +0000 | [diff] [blame] | 7754 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7755 | #define _mm_getmant_pd(A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7756 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 7757 | (int)(((C)<<2) | (B)), \ |
| 7758 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7759 | (__mmask8)-1) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7760 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7761 | #define _mm_mask_getmant_pd(W, U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7762 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 7763 | (int)(((C)<<2) | (B)), \ |
| 7764 | (__v2df)(__m128d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7765 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7766 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7767 | #define _mm_maskz_getmant_pd(U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7768 | (__m128d)__builtin_ia32_getmantpd128_mask((__v2df)(__m128d)(A), \ |
| 7769 | (int)(((C)<<2) | (B)), \ |
| 7770 | (__v2df)_mm_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7771 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7772 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7773 | #define _mm256_getmant_pd(A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7774 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 7775 | (int)(((C)<<2) | (B)), \ |
| 7776 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7777 | (__mmask8)-1) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7778 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7779 | #define _mm256_mask_getmant_pd(W, U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7780 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 7781 | (int)(((C)<<2) | (B)), \ |
| 7782 | (__v4df)(__m256d)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7783 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7784 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7785 | #define _mm256_maskz_getmant_pd(U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7786 | (__m256d)__builtin_ia32_getmantpd256_mask((__v4df)(__m256d)(A), \ |
| 7787 | (int)(((C)<<2) | (B)), \ |
| 7788 | (__v4df)_mm256_setzero_pd(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7789 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7790 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7791 | #define _mm_getmant_ps(A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7792 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 7793 | (int)(((C)<<2) | (B)), \ |
| 7794 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7795 | (__mmask8)-1) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7796 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7797 | #define _mm_mask_getmant_ps(W, U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7798 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 7799 | (int)(((C)<<2) | (B)), \ |
| 7800 | (__v4sf)(__m128)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7801 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7802 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7803 | #define _mm_maskz_getmant_ps(U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7804 | (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \ |
| 7805 | (int)(((C)<<2) | (B)), \ |
| 7806 | (__v4sf)_mm_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7807 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7808 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7809 | #define _mm256_getmant_ps(A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7810 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 7811 | (int)(((C)<<2) | (B)), \ |
| 7812 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7813 | (__mmask8)-1) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7814 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7815 | #define _mm256_mask_getmant_ps(W, U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7816 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 7817 | (int)(((C)<<2) | (B)), \ |
| 7818 | (__v8sf)(__m256)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7819 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7820 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7821 | #define _mm256_maskz_getmant_ps(U, A, B, C) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7822 | (__m256)__builtin_ia32_getmantps256_mask((__v8sf)(__m256)(A), \ |
| 7823 | (int)(((C)<<2) | (B)), \ |
| 7824 | (__v8sf)_mm256_setzero_ps(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7825 | (__mmask8)(U)) |
Michael Zuckerman | 6fa512c | 2016-04-19 17:10:29 +0000 | [diff] [blame] | 7826 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7827 | #define _mm_mmask_i64gather_pd(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7828 | (__m128d)__builtin_ia32_gather3div2df((__v2df)(__m128d)(v1_old), \ |
| 7829 | (double const *)(addr), \ |
| 7830 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7831 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7832 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7833 | #define _mm_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7834 | (__m128i)__builtin_ia32_gather3div2di((__v2di)(__m128i)(v1_old), \ |
| 7835 | (long long const *)(addr), \ |
| 7836 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7837 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7838 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7839 | #define _mm256_mmask_i64gather_pd(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7840 | (__m256d)__builtin_ia32_gather3div4df((__v4df)(__m256d)(v1_old), \ |
| 7841 | (double const *)(addr), \ |
| 7842 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7843 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7844 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7845 | #define _mm256_mmask_i64gather_epi64(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7846 | (__m256i)__builtin_ia32_gather3div4di((__v4di)(__m256i)(v1_old), \ |
| 7847 | (long long const *)(addr), \ |
| 7848 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7849 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7850 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7851 | #define _mm_mmask_i64gather_ps(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7852 | (__m128)__builtin_ia32_gather3div4sf((__v4sf)(__m128)(v1_old), \ |
| 7853 | (float const *)(addr), \ |
| 7854 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7855 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7856 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7857 | #define _mm_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7858 | (__m128i)__builtin_ia32_gather3div4si((__v4si)(__m128i)(v1_old), \ |
| 7859 | (int const *)(addr), \ |
| 7860 | (__v2di)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7861 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7862 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7863 | #define _mm256_mmask_i64gather_ps(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7864 | (__m128)__builtin_ia32_gather3div8sf((__v4sf)(__m128)(v1_old), \ |
| 7865 | (float const *)(addr), \ |
| 7866 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7867 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7868 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7869 | #define _mm256_mmask_i64gather_epi32(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7870 | (__m128i)__builtin_ia32_gather3div8si((__v4si)(__m128i)(v1_old), \ |
| 7871 | (int const *)(addr), \ |
| 7872 | (__v4di)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7873 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7874 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7875 | #define _mm_mmask_i32gather_pd(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7876 | (__m128d)__builtin_ia32_gather3siv2df((__v2df)(__m128d)(v1_old), \ |
| 7877 | (double const *)(addr), \ |
| 7878 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7879 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7880 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7881 | #define _mm_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7882 | (__m128i)__builtin_ia32_gather3siv2di((__v2di)(__m128i)(v1_old), \ |
| 7883 | (long long const *)(addr), \ |
| 7884 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7885 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7886 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7887 | #define _mm256_mmask_i32gather_pd(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7888 | (__m256d)__builtin_ia32_gather3siv4df((__v4df)(__m256d)(v1_old), \ |
| 7889 | (double const *)(addr), \ |
| 7890 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7891 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7892 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7893 | #define _mm256_mmask_i32gather_epi64(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7894 | (__m256i)__builtin_ia32_gather3siv4di((__v4di)(__m256i)(v1_old), \ |
| 7895 | (long long const *)(addr), \ |
| 7896 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7897 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7898 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7899 | #define _mm_mmask_i32gather_ps(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7900 | (__m128)__builtin_ia32_gather3siv4sf((__v4sf)(__m128)(v1_old), \ |
| 7901 | (float const *)(addr), \ |
| 7902 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7903 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7904 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7905 | #define _mm_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7906 | (__m128i)__builtin_ia32_gather3siv4si((__v4si)(__m128i)(v1_old), \ |
| 7907 | (int const *)(addr), \ |
| 7908 | (__v4si)(__m128i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7909 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7910 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7911 | #define _mm256_mmask_i32gather_ps(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7912 | (__m256)__builtin_ia32_gather3siv8sf((__v8sf)(__m256)(v1_old), \ |
| 7913 | (float const *)(addr), \ |
| 7914 | (__v8si)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7915 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7916 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7917 | #define _mm256_mmask_i32gather_epi32(v1_old, mask, index, addr, scale) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 7918 | (__m256i)__builtin_ia32_gather3siv8si((__v8si)(__m256i)(v1_old), \ |
| 7919 | (int const *)(addr), \ |
| 7920 | (__v8si)(__m256i)(index), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7921 | (__mmask8)(mask), (int)(scale)) |
Michael Zuckerman | 4fa96af | 2016-04-21 12:47:27 +0000 | [diff] [blame] | 7922 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7923 | #define _mm256_permutex_pd(X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7924 | (__m256d)__builtin_shufflevector((__v4df)(__m256d)(X), \ |
| 7925 | (__v4df)_mm256_undefined_pd(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 7926 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7927 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7928 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7929 | #define _mm256_mask_permutex_pd(W, U, X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7930 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7931 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7932 | (__v4df)(__m256d)(W)) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7933 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7934 | #define _mm256_maskz_permutex_pd(U, X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7935 | (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 7936 | (__v4df)_mm256_permutex_pd((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7937 | (__v4df)_mm256_setzero_pd()) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7938 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7939 | #define _mm256_permutex_epi64(X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7940 | (__m256i)__builtin_shufflevector((__v4di)(__m256i)(X), \ |
| 7941 | (__v4di)_mm256_undefined_si256(), \ |
Craig Topper | 2a383c9 | 2016-07-04 22:18:01 +0000 | [diff] [blame] | 7942 | ((C) >> 0) & 0x3, ((C) >> 2) & 0x3, \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7943 | ((C) >> 4) & 0x3, ((C) >> 6) & 0x3) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7944 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7945 | #define _mm256_mask_permutex_epi64(W, U, X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7946 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 7947 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7948 | (__v4di)(__m256i)(W)) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7949 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7950 | #define _mm256_maskz_permutex_epi64(U, X, C) \ |
Simon Pilgrim | 30db811 | 2016-07-04 13:34:44 +0000 | [diff] [blame] | 7951 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 7952 | (__v4di)_mm256_permutex_epi64((X), (C)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 7953 | (__v4di)_mm256_setzero_si256()) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7954 | |
| 7955 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7956 | _mm256_permutexvar_pd (__m256i __X, __m256d __Y) |
| 7957 | { |
Craig Topper | 55b4067 | 2018-05-20 23:34:10 +0000 | [diff] [blame] | 7958 | return (__m256d)__builtin_ia32_permvardf256((__v4df)__Y, (__v4di)__X); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7959 | } |
| 7960 | |
| 7961 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7962 | _mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X, |
| 7963 | __m256d __Y) |
| 7964 | { |
Craig Topper | 55b4067 | 2018-05-20 23:34:10 +0000 | [diff] [blame] | 7965 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 7966 | (__v4df)_mm256_permutexvar_pd(__X, __Y), |
| 7967 | (__v4df)__W); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7968 | } |
| 7969 | |
| 7970 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 7971 | _mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y) |
| 7972 | { |
Craig Topper | 55b4067 | 2018-05-20 23:34:10 +0000 | [diff] [blame] | 7973 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 7974 | (__v4df)_mm256_permutexvar_pd(__X, __Y), |
| 7975 | (__v4df)_mm256_setzero_pd()); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7976 | } |
| 7977 | |
| 7978 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Michael Zuckerman | a72b49ef | 2016-07-05 11:30:31 +0000 | [diff] [blame] | 7979 | _mm256_permutexvar_epi64 ( __m256i __X, __m256i __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7980 | { |
Craig Topper | 55b4067 | 2018-05-20 23:34:10 +0000 | [diff] [blame] | 7981 | return (__m256i)__builtin_ia32_permvardi256((__v4di) __Y, (__v4di) __X); |
| 7982 | } |
| 7983 | |
| 7984 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7985 | _mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y) |
| 7986 | { |
| 7987 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 7988 | (__v4di)_mm256_permutexvar_epi64(__X, __Y), |
| 7989 | (__v4di)_mm256_setzero_si256()); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7990 | } |
| 7991 | |
| 7992 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 7993 | _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X, |
| 7994 | __m256i __Y) |
| 7995 | { |
Craig Topper | 55b4067 | 2018-05-20 23:34:10 +0000 | [diff] [blame] | 7996 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 7997 | (__v4di)_mm256_permutexvar_epi64(__X, __Y), |
| 7998 | (__v4di)__W); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 7999 | } |
| 8000 | |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8001 | #define _mm256_permutexvar_ps(A, B) _mm256_permutevar8x32_ps((B), (A)) |
| 8002 | |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8003 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8004 | _mm256_mask_permutexvar_ps(__m256 __W, __mmask8 __U, __m256i __X, __m256 __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8005 | { |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8006 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8007 | (__v8sf)_mm256_permutexvar_ps(__X, __Y), |
| 8008 | (__v8sf)__W); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8009 | } |
| 8010 | |
| 8011 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8012 | _mm256_maskz_permutexvar_ps(__mmask8 __U, __m256i __X, __m256 __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8013 | { |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8014 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8015 | (__v8sf)_mm256_permutexvar_ps(__X, __Y), |
| 8016 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8017 | } |
| 8018 | |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8019 | #define _mm256_permutexvar_epi32(A, B) _mm256_permutevar8x32_epi32((B), (A)) |
| 8020 | |
| 8021 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
| 8022 | _mm256_mask_permutexvar_epi32(__m256i __W, __mmask8 __M, __m256i __X, |
| 8023 | __m256i __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8024 | { |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8025 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 8026 | (__v8si)_mm256_permutexvar_epi32(__X, __Y), |
| 8027 | (__v8si)__W); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8028 | } |
| 8029 | |
| 8030 | static __inline__ __m256i __DEFAULT_FN_ATTRS |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8031 | _mm256_maskz_permutexvar_epi32(__mmask8 __M, __m256i __X, __m256i __Y) |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8032 | { |
Craig Topper | 5dc6ca8 | 2018-02-24 06:46:42 +0000 | [diff] [blame] | 8033 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 8034 | (__v8si)_mm256_permutexvar_epi32(__X, __Y), |
| 8035 | (__v8si)_mm256_setzero_si256()); |
Michael Zuckerman | 8938e83 | 2016-04-25 05:32:35 +0000 | [diff] [blame] | 8036 | } |
| 8037 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8038 | #define _mm_alignr_epi32(A, B, imm) \ |
Craig Topper | e56819e | 2018-06-07 21:27:41 +0000 | [diff] [blame] | 8039 | (__m128i)__builtin_ia32_alignd128((__v4si)(__m128i)(A), \ |
| 8040 | (__v4si)(__m128i)(B), (int)(imm)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8041 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8042 | #define _mm_mask_alignr_epi32(W, U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8043 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8044 | (__v4si)_mm_alignr_epi32((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8045 | (__v4si)(__m128i)(W)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8046 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8047 | #define _mm_maskz_alignr_epi32(U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8048 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8049 | (__v4si)_mm_alignr_epi32((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8050 | (__v4si)_mm_setzero_si128()) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8051 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8052 | #define _mm256_alignr_epi32(A, B, imm) \ |
Craig Topper | e56819e | 2018-06-07 21:27:41 +0000 | [diff] [blame] | 8053 | (__m256i)__builtin_ia32_alignd256((__v8si)(__m256i)(A), \ |
| 8054 | (__v8si)(__m256i)(B), (int)(imm)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8055 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8056 | #define _mm256_mask_alignr_epi32(W, U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8057 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8058 | (__v8si)_mm256_alignr_epi32((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8059 | (__v8si)(__m256i)(W)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8060 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8061 | #define _mm256_maskz_alignr_epi32(U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8062 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8063 | (__v8si)_mm256_alignr_epi32((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8064 | (__v8si)_mm256_setzero_si256()) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8065 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8066 | #define _mm_alignr_epi64(A, B, imm) \ |
Craig Topper | e56819e | 2018-06-07 21:27:41 +0000 | [diff] [blame] | 8067 | (__m128i)__builtin_ia32_alignq128((__v2di)(__m128i)(A), \ |
| 8068 | (__v2di)(__m128i)(B), (int)(imm)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8069 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8070 | #define _mm_mask_alignr_epi64(W, U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8071 | (__m128i)__builtin_ia32_selectq_128((__mmask8)(U), \ |
| 8072 | (__v2di)_mm_alignr_epi64((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8073 | (__v2di)(__m128i)(W)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8074 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8075 | #define _mm_maskz_alignr_epi64(U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8076 | (__m128i)__builtin_ia32_selectq_128((__mmask8)(U), \ |
| 8077 | (__v2di)_mm_alignr_epi64((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8078 | (__v2di)_mm_setzero_si128()) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8079 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8080 | #define _mm256_alignr_epi64(A, B, imm) \ |
Craig Topper | e56819e | 2018-06-07 21:27:41 +0000 | [diff] [blame] | 8081 | (__m256i)__builtin_ia32_alignq256((__v4di)(__m256i)(A), \ |
| 8082 | (__v4di)(__m256i)(B), (int)(imm)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8083 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8084 | #define _mm256_mask_alignr_epi64(W, U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8085 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8086 | (__v4di)_mm256_alignr_epi64((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8087 | (__v4di)(__m256i)(W)) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8088 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8089 | #define _mm256_maskz_alignr_epi64(U, A, B, imm) \ |
Craig Topper | 6aefe00 | 2016-11-23 01:47:12 +0000 | [diff] [blame] | 8090 | (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 8091 | (__v4di)_mm256_alignr_epi64((A), (B), (imm)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8092 | (__v4di)_mm256_setzero_si256()) |
Michael Zuckerman | 533e065 | 2016-04-28 12:47:30 +0000 | [diff] [blame] | 8093 | |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8094 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8095 | _mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8096 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8097 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8098 | (__v4sf)_mm_movehdup_ps(__A), |
| 8099 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8100 | } |
| 8101 | |
| 8102 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8103 | _mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A) |
| 8104 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8105 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8106 | (__v4sf)_mm_movehdup_ps(__A), |
| 8107 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8108 | } |
| 8109 | |
| 8110 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8111 | _mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8112 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8113 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8114 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8115 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8116 | } |
| 8117 | |
| 8118 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8119 | _mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A) |
| 8120 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8121 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8122 | (__v8sf)_mm256_movehdup_ps(__A), |
| 8123 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8124 | } |
| 8125 | |
| 8126 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8127 | _mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8128 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8129 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8130 | (__v4sf)_mm_moveldup_ps(__A), |
| 8131 | (__v4sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8132 | } |
| 8133 | |
| 8134 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8135 | _mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A) |
| 8136 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8137 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 8138 | (__v4sf)_mm_moveldup_ps(__A), |
| 8139 | (__v4sf)_mm_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8140 | } |
| 8141 | |
| 8142 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8143 | _mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8144 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8145 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8146 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8147 | (__v8sf)__W); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8148 | } |
| 8149 | |
| 8150 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8151 | _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A) |
| 8152 | { |
Simon Pilgrim | 275d721 | 2016-07-02 17:16:25 +0000 | [diff] [blame] | 8153 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 8154 | (__v8sf)_mm256_moveldup_ps(__A), |
| 8155 | (__v8sf)_mm256_setzero_ps()); |
Michael Zuckerman | ac1e519 | 2016-05-01 14:43:43 +0000 | [diff] [blame] | 8156 | } |
| 8157 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8158 | #define _mm256_mask_shuffle_epi32(W, U, A, I) \ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8159 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8160 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8161 | (__v8si)(__m256i)(W)) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8162 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8163 | #define _mm256_maskz_shuffle_epi32(U, A, I) \ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8164 | (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \ |
| 8165 | (__v8si)_mm256_shuffle_epi32((A), (I)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8166 | (__v8si)_mm256_setzero_si256()) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8167 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8168 | #define _mm_mask_shuffle_epi32(W, U, A, I) \ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8169 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8170 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8171 | (__v4si)(__m128i)(W)) |
Michael Zuckerman | c62f27e | 2016-05-02 07:35:27 +0000 | [diff] [blame] | 8172 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8173 | #define _mm_maskz_shuffle_epi32(U, A, I) \ |
Craig Topper | 7cc9263 | 2016-06-11 12:50:19 +0000 | [diff] [blame] | 8174 | (__m128i)__builtin_ia32_selectd_128((__mmask8)(U), \ |
| 8175 | (__v4si)_mm_shuffle_epi32((A), (I)), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8176 | (__v4si)_mm_setzero_si128()) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8177 | |
| 8178 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 8179 | _mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A) |
| 8180 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8181 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 8182 | (__v2df) __A, |
| 8183 | (__v2df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8184 | } |
| 8185 | |
| 8186 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
| 8187 | _mm_maskz_mov_pd (__mmask8 __U, __m128d __A) |
| 8188 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8189 | return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U, |
| 8190 | (__v2df) __A, |
| 8191 | (__v2df) _mm_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8192 | } |
| 8193 | |
| 8194 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8195 | _mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A) |
| 8196 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8197 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 8198 | (__v4df) __A, |
| 8199 | (__v4df) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8200 | } |
| 8201 | |
| 8202 | static __inline__ __m256d __DEFAULT_FN_ATTRS |
| 8203 | _mm256_maskz_mov_pd (__mmask8 __U, __m256d __A) |
| 8204 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8205 | return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U, |
| 8206 | (__v4df) __A, |
| 8207 | (__v4df) _mm256_setzero_pd ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8208 | } |
| 8209 | |
| 8210 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8211 | _mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A) |
| 8212 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8213 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 8214 | (__v4sf) __A, |
| 8215 | (__v4sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8216 | } |
| 8217 | |
| 8218 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8219 | _mm_maskz_mov_ps (__mmask8 __U, __m128 __A) |
| 8220 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8221 | return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U, |
| 8222 | (__v4sf) __A, |
| 8223 | (__v4sf) _mm_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8224 | } |
| 8225 | |
| 8226 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8227 | _mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A) |
| 8228 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8229 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 8230 | (__v8sf) __A, |
| 8231 | (__v8sf) __W); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8232 | } |
| 8233 | |
| 8234 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8235 | _mm256_maskz_mov_ps (__mmask8 __U, __m256 __A) |
| 8236 | { |
Igor Breger | aadb876 | 2016-06-08 13:59:20 +0000 | [diff] [blame] | 8237 | return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U, |
| 8238 | (__v8sf) __A, |
| 8239 | (__v8sf) _mm256_setzero_ps ()); |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8240 | } |
| 8241 | |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 8242 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8243 | _mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A) |
| 8244 | { |
| 8245 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 8246 | (__v4sf) __W, |
| 8247 | (__mmask8) __U); |
| 8248 | } |
| 8249 | |
| 8250 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
| 8251 | _mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 8252 | { |
| 8253 | return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, |
| 8254 | (__v4sf) |
| 8255 | _mm_setzero_ps (), |
| 8256 | (__mmask8) __U); |
| 8257 | } |
| 8258 | |
| 8259 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8260 | _mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A) |
| 8261 | { |
| 8262 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 8263 | (__v8sf) __W, |
| 8264 | (__mmask8) __U); |
| 8265 | } |
| 8266 | |
| 8267 | static __inline__ __m256 __DEFAULT_FN_ATTRS |
| 8268 | _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A) |
| 8269 | { |
| 8270 | return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, |
| 8271 | (__v8sf) |
| 8272 | _mm256_setzero_ps (), |
| 8273 | (__mmask8) __U); |
| 8274 | } |
| 8275 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 8276 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 8277 | _mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A) |
| 8278 | { |
| 8279 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 8280 | (__v8hi) __W, |
| 8281 | (__mmask8) __U); |
| 8282 | } |
| 8283 | |
| 8284 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 8285 | _mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A) |
| 8286 | { |
| 8287 | return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 8288 | (__v8hi) _mm_setzero_si128 (), |
| 8289 | (__mmask8) __U); |
| 8290 | } |
| 8291 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8292 | #define _mm_mask_cvt_roundps_ph(W, U, A, I) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8293 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 8294 | (__v8hi)(__m128i)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8295 | (__mmask8)(U)) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 8296 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8297 | #define _mm_maskz_cvt_roundps_ph(U, A, I) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8298 | (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \ |
| 8299 | (__v8hi)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8300 | (__mmask8)(U)) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 8301 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 8302 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 8303 | _mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A) |
| 8304 | { |
| 8305 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 8306 | (__v8hi) __W, |
| 8307 | (__mmask8) __U); |
| 8308 | } |
| 8309 | |
| 8310 | static __inline __m128i __DEFAULT_FN_ATTRS |
| 8311 | _mm256_maskz_cvtps_ph ( __mmask8 __U, __m256 __A) |
| 8312 | { |
| 8313 | return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, _MM_FROUND_CUR_DIRECTION, |
| 8314 | (__v8hi) _mm_setzero_si128(), |
| 8315 | (__mmask8) __U); |
| 8316 | } |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8317 | #define _mm256_mask_cvt_roundps_ph(W, U, A, I) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8318 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 8319 | (__v8hi)(__m128i)(W), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8320 | (__mmask8)(U)) |
Michael Zuckerman | 8bfb777 | 2016-05-03 12:45:04 +0000 | [diff] [blame] | 8321 | |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8322 | #define _mm256_maskz_cvt_roundps_ph(U, A, I) \ |
Craig Topper | d266188 | 2016-05-17 04:41:48 +0000 | [diff] [blame] | 8323 | (__m128i)__builtin_ia32_vcvtps2ph256_mask((__v8sf)(__m256)(A), (int)(I), \ |
| 8324 | (__v8hi)_mm_setzero_si128(), \ |
Craig Topper | c633867 | 2018-05-31 00:51:20 +0000 | [diff] [blame] | 8325 | (__mmask8)(U)) |
Michael Zuckerman | 5f0e96e | 2016-05-02 14:02:01 +0000 | [diff] [blame] | 8326 | |
Michael Zuckerman | 9e7d0a9 | 2016-06-02 07:44:08 +0000 | [diff] [blame] | 8327 | |
Michael Kuperstein | e45af54 | 2015-06-30 13:36:19 +0000 | [diff] [blame] | 8328 | #undef __DEFAULT_FN_ATTRS |
Eric Christopher | 4d185168 | 2015-06-17 07:09:20 +0000 | [diff] [blame] | 8329 | |
Robert Khasanov | b9f3a91 | 2014-10-08 17:18:13 +0000 | [diff] [blame] | 8330 | #endif /* __AVX512VLINTRIN_H */ |